Can we live with AI? — Shannon Vallor (Turing Lecture, Dec 2024)

Shannon Vallor · 2024 · 1:33:16 · ▶ Watch

Summary

Central question. “Can we live with AI?” She splits it three ways. Survive with AI (most people in the room can). Live well with it (is it minimally acceptable, does it make life lesser?). And the word can: possibility, not probability. We should want confidence that we will live well with it, not only that we could.

Main thesis.

1. Today’s AI (machine learning on large human data, especially large language models) is best understood as a mirror, not a mind. Its outputs are “composite reflections” of data taken from us. A reflection of a mind is not a mind.
2. Used badly, the mirror steals three things from us: space (room to think), time (an open future), and stories (our power to author our own lives).
3. AI is built “pointing backward”. It cannot step outside the pattern of past data, so it cannot do what Ortega called autofabrication, the human power to choose a new direction.
4. The market wants productivity with fewer steps, which shrinks human thinking space. The result is a “mental space filling machine”.

Close. “we can live with AI if we choose to but we cannot live through it”. Humans must keep the right to be the authors.

Stance to note. She is not anti-AI. She wants “the kind of AI that can tell me what part of my city will flood tomorrow”. She names the target as the political economy and the narrative (innovation without progress), not the maths. In Q&A she says real existential risk lies in energy demand and in the loss of an open future, not in machines “waking up”.

Map — talking points

#	Point
1	Welcome (Knowledge Quarter), series intro (Turing Lectures since 2016)
2	Slido poll: what could AI never replace? Can you live without AI?
3	Who can live with AI? Gaza, authoritarian use, thirst from data-centre water, Three Mile Island, Omaha
4	Eric Schmidt clip: bet on AI to solve the climate problem; “needs” of the technology
5	Three readings of “can we live with AI?” (survive, live well, can vs will)
6	”We used to get excited about technology” (MIT Tech Review, 2022). What we wanted vs what we got
7	Mirror metaphor: why a mirror beats a mind. Three components (polished surface, light, reflection)
8	A reflection of a mind is not a mind. Narcissus. The mirror test
9	Distortion: bias magnified, funhouse mirrors, absurd errors, deepfakes. Narrow models that reveal (cancer, sepsis)
10	Three things stolen: space, time, stories. Sellars’ space of reasons; open mind vs closed
11	AI as rear-view mirror; “stories never written” test; Goldman Sachs; Whitehead; the catapult-robot and mountain
12	Thinking as a skill; the market wants fewer steps
13	Time: Ricoeur on narrative; light cone of the past; temperature is a wider guess, not an open future
14	Husserl’s horizon; Hume; Assad; Ortega’s autofabrication; Coeckelbergh’s “time machine”
15	We need to steer; AI as “thieves’ tools”; the attention trap and unfilled time
16	Stories: journalism, fiction, summaries of your own life, education, UCLA course, students who “can’t read books”
17	Boredom as a portal; the guitar, rock, Bible; Tetris metaphor for modern life
18	Coeckelbergh quote; no story untold for a mirror; the closing line
19	Q&A: how to choose which AI without exploring useless ones? AlphaFold vs forced features
20	Q&A: role of governments, arbitrage, public-good AI
21	Q&A: “lost the plot”: innovation vs progress; get constructively mad
22	Q&A: levers; academia, media coziness, “what kind of AI do we deserve”
23	Q&A: democracy and marginalised groups; “master’s tools”
24	Q&A: existential risk; Schmidt “not entirely wrong”; efficiency rebound; the philosophical existential risk; AI harms are people’s harms

Putting it together

The mental model in one paragraph. Today’s AI is a mirror made of maths and our own data. Because the data come from thinking, feeling people, the reflections look like thought. They are useful where the pattern in the past is what we need (a tumour, a protein), and harmful where the market uses them to fill every gap in our thinking, time and storytelling with plausible answers. Humans live in an open space of reasons and an open future, and we build our selves by climbing, choosing and telling. A mirror cannot choose a new direction, so letting it steer means giving the future to the past. The answer to “can we live with AI?” is “yes, if we keep authorship”; the political task is to change the incentives that reward the opposite.

Three tensions to keep open

1. Mirror or more? Interpretability evidence for internal world models vs. her claim that outputs are bounded by past data. Both sides agree models are not minds; they disagree on how much structure and novelty are inside.
2. Productivity vs practice. Is offloading the climb a loss, or the same trade we made with writing and calculators? The answer depends on which skills are constitutive of being a thinker.
3. Moral risk vs material risk. Her “mindless, so risk comes from people” view vs. the safety view that capable optimisers need guarding whether or not they have minds.

Reading path

• Start: Vallor, The AI Mirror (2024), introduction and chapter 1. Her Noema essay “What Thoughts Do the Civilized Keep?” (2021).
• Then: Sellars, “Empiricism and the Philosophy of Mind” (read the Introduction and Part VIII with a guide). Ricoeur, Time and Narrative vol. 1, ch. 1-3. Bender et al., “Stochastic Parrots” (2021). Bainbridge, “Ironies of Automation” (1983, short).
• Deeper: Vallor, Technology and the Virtues (2016). Coeckelbergh, Time Machines (2021). Jonas, The Imperative of Responsibility. Winner, “Do Artifacts Have Politics?“. Crawford, Atlas of AI. Benjamin, “The Storyteller”. The Othello-GPT paper and the Millière-Buckner survey for the technical edge.

Part 1 — Before you read

Key terms

• AI mirror. Vallor’s metaphor and book title (The AI Mirror, OUP, 2024). A machine-learning model is a “polished” mathematical surface; training data is the “light”; outputs are “reflections”. She restricts it to machine learning on large human-made data. It fits narrow models (cancer detection) as well as LLMs.
• Space of reasons. A phrase from Wilfrid Sellars (1956): the social realm where claims are justified and challenged. She uses it spatially: thinking needs room to move.
• Light cone. A physics image: the region of events that can influence a point. She uses it for the idea that AI sees only what is already in the past of the data.
• Temperature. A model setting that makes output more random. She argues it widens the guess within past patterns; it does not leave them.
• Autofabrication (Ortega y Gasset): self-making. A human is “an engineer of yourself”.
• Temporal horizon (Husserl): the open, indeterminate edge of experienced time.
• “Mindless”: her label for machines with no understanding, wants or stakes. Compare with “mind” in the AI industry, where it means a system that performs mental tasks. Watch out: these two senses are what the whole talk turns on.
• Innovation vs progress. Innovation is any new thing. Progress is a new thing that makes life better.
• Existential risk, two senses. Common: risk that humanity ends. Vallor’s: risk to the human capacity to keep the future open. See idea 6.

People named in the talk (a reading list)

Person	Why they come up	Start with
Eric Schmidt	Video clip: bet on AI to solve climate; “alien intelligence”	Clip
Wilfrid Sellars	”Space of reasons"	"Empiricism and the Philosophy of Mind” (1956)
Socrates	Questions hold the space of reasons open	Plato’s early dialogues
Alfred North Whitehead	Civilisation advances by operations we need not think about	An Introduction to Mathematics (1911)
Paul Ricoeur (likely)	Narrative makes selves and time	Time and Narrative (1983-85)
Edmund Husserl	Temporal horizon	Lectures on the Phenomenology of Internal Time-Consciousness
David Hume	Induction: no evidence the future will match the past	Enquiry section IV
José Ortega y Gasset	Autofabrication	History as a System, Meditations on Technics
Mark Coeckelbergh (likely)	AI as a time machine of the past; make time for the future	Time Machines (2021)
Abeba Birhane (low confidence)	“AI is an inherently conservative technology”	Her essays on algorithmic injustice
Rose Horowitch	Atlantic piece: “The Elite College Students Who Can’t Read Books”	The Atlantic, 2024

Vallor’s own frame

Vallor is a virtue ethicist of technology. Her first book, Technology and the Virtues (OUP, 2016), argues for “technomoral virtues” (humility, honesty, care, courage and so on) as the right vocabulary for living with fast tech change. The AI Mirror applies this to AI. The 2022 MIT Technology Review essay she mentions is “We used to get excited about technology. What happened?” and her Noema essay “What Thoughts Do the Civilized Keep?” (2021) is the source of the catapult image. Outside this lecture, she has also been a senior AI ethics voice at Google (2018-20)†.

The current debate (2024-26)

• Reception of the mirror thesis. The book has had many academic reviews. Reviewers praise the move from “machine agency” to “human self-formation” (BJPS Review of Books, PDF; Technology and Society review; Critical Inquiry; The Scotsman). A search summary reports one reviewer, Sarah Richmond, arguing that Vallor overlooks existing policy work to regulate AI and Big Tech’s role in blocking it; I have not read that review itself. I did not find a systematic technical rebuttal in the results.
• Is “mirror” technically right? This is the live question. The same-period interpretability work on world models (for example the Othello-GPT study, Li et al., arXiv) reports that a model trained only on move sequences builds an internal board representation that causally affects its output. That fits “reflection” only loosely: the model is not copying surface text, it is building structure. A philosophy survey of the debate concludes that behavioural tests cannot settle it and that mechanistic interpretability looks promising (Millière and Buckner, Part II, arXiv). Her mirror claim and the world-model finding can both be partly true: structure learned from the past, but still bound to the data’s distribution. Whether it can reach beyond that is open. The search I ran did not address novelty directly.
• Energy. After the lecture, coverage of the IEA’s April 2026 Key Questions on Energy and AI reports data-centre electricity use up about 17% in 2025 and expected to roughly double by 2030 (about 950 TWh, around 3% of global electricity), with efficiency gains offset by rising use, a Jevons-style rebound (Rigzone; PV Tech). I read only secondary reports, not the IEA document. This supports her Q&A point that efficiency will not cut total demand. It does not settle Schmidt’s other bet, that AI will pay the climate bill back.
• Cognitive offloading. The best-known evidence for “thinking atrophies” is MIT Media Lab’s Your Brain on ChatGPT (2025; small sample, preprint): MIT Media Lab. It is suggestive, not proof of her catapult claim.

The ideas

Vallor spoke for 45 minutes to a general audience, so the philosophy is mostly implicit. Eight families of ideas meet in the talk.

 MIRRORS & AI CRITIQUE ───────┐                  ┌──── SPACE OF REASONS
 Ovid · Rorty · Weizenbaum    │                  │     Sellars · McDowell · Brandom
 Bender · Searle · Dreyfus    ▼                  ▼     Socrates · Kant · Habermas
                      ┌──────────────────────────────────┐
 TIME & NARRATIVE ──► │  VALLOR: AI as a mirror points   │ ◄── DESKILLING & OFFLOADING
 Husserl · Hume ·     │  backward and fills the space,   │     Plato · Whitehead · Carr
 Ricoeur · Ortega     │  time and stories we need to be  │     Bainbridge · Stiegler
                      │  free                            │
                      └──────────────────────────────────┘
                        ▲            ▲             ▲
 ATTENTION & BOREDOM    │  VIRTUE ETHICS OF TECH  │  POLITICAL ECONOMY & ENERGY
 Pascal · Simon · Wu    │  Aristotle · MacIntyre  │  Jevons · Winner · Zuboff
 Winnicott · Rosa       │  Vallor · Borgmann      │  Crawford · Mazzucato
                        │                         │
                    HOW MODELS WORK: next-token · distribution shift · world models

Where each school of thought is covered:

• Mirrors and the critique of “machine mind”: idea 1
• The space of reasons: idea 2
• Time, narrative and the open future: idea 3
• Offloading and deskilling: idea 2
• Attention, boredom and unfilled space: idea 4
• Political economy, energy and “innovation”: idea 5
• Virtue ethics of technology: idea 6
• How the models actually work: idea 3

1. The mirror: why “reflection” and not “mind”

In short. Vallor says the mirror is a far more apt, illuminating and safer picture of AI than the familiar picture of a mind, and far less deceptive. All technology mirrors human values, but today’s AI tools are mirrors in a much more powerful sense. She limits the claim: AI is not one technology but many, and the mirror fits best machine learning tools trained on large volumes of human-generated data.

🎙 In the talk. She says “the mirror metaphor for AI is far more apt and Illuminating and much safer and less deceptive than the familiar metaphor” of a mind. All technology is a mirror of human values, but “today’s AI tools are mirrors in a much more powerful sense”. She limits the claim: “AI is not one technology it’s many”, and the metaphor is “best suited to describe machine learning tools trained on large volumes of human generated data”.

The three components: the learning algorithm is the polished surface; data is the light (“all the hoarded data that large AI companies have grabbed and scraped and sometimes stolen from us”); the outputs are “composite reflections of the structures of the data taken from us”.

Why they seem alive: “because they’re reflections of data taken from intelligent conscious living beings from us”. Then the key line: “a reflection of a mind is not a mind and it does not think”. She says AI leaders “very much want you to fail the mirror test”.

Mirrors do three things: reveal (cancer, sepsis, insider trading from narrow models), magnify (biases and stereotypes), distort (funhouse images, deepfakes, fraud). She is grateful for the silly errors because they “puncture the mirror illusion”. She adds that correct outputs involve “no more thinking Behind the Reflection”.

📚 Deeper theory

• Links the Roots of this idea (Ovid, Turing, Weizenbaum, Searle, Bender) and idea 3 (distribution shift).
• A mirror is passive; a model is not. A glass mirror cannot be “fine-tuned” by its light in the way she imagines, and she says so, using it as a thought experiment. A better objection: a model recombines patterns across millions of sources, so its output can be a new combination. Vallor’s answer is that new combinations of old patterns are still within the data cone (see idea 3).
• Why “mind” is not harmless. A metaphor chooses the questions you ask. “Mind” invites questions of rights, alignment, and rivalry. “Mirror” invites questions of data provenance, bias, and who holds the glass. The choice is also political (compare Gebru in gebru-machine-god-doom-as-marketing).
• Reveal and magnify are legitimate. She is careful that mirrors are useful. This puts her closer to the cautious pragmatist than the dismissive “stochastic parrot” camp.

📚 Roots: Mirrors and the critique of “machine mind”

The problem. When a system talks like a person, people credit it with a mind. Is that a fair inference, or a trick of the surface?

The point. The mirror gives her a way to say that the outputs are about us (our data) and not about a new someone. It is a move from “is it a mind?” to “what is it a picture of?“.

1. Ovid, Metamorphoses (c. 8 CE). Narcissus dies in love with his reflection. The point Vallor uses: the danger is not the mirror but mistaking it for a second person. (She says “Roman poet”, which fits Ovid.)
2. Alan Turing, “Computing Machinery and Intelligence” (1950). The lecture series bears his name. He replaced “can machines think?” with the imitation game: judge by conversational behaviour. Vallor’s mirror argument is a direct reply to this move: behaviour that looks intelligent can be reflection.
3. Joseph Weizenbaum, ELIZA (1966) and Computer Power and Human Reason (1976). A simple script made users confide in it. He was disturbed that people attributed understanding to it. The “ELIZA effect” is the name for what Vallor describes. He also argued that some tasks should not be handed to machines, even if they could do them. That is close to her “what AI should not do” theme.
4. John Searle, “Minds, Brains, and Programs” (1980). The Chinese Room: symbol manipulation is not understanding. Vallor shares the conclusion for LLMs but argues from the data-reflection structure, not from syntax alone (from memory; she does not name Searle).
5. Richard Rorty, Philosophy and the Mirror of Nature (1979). Rorty attacked the idea that the mind mirrors reality. Different target (knowledge, not AI), but the irony is useful: in Rorty’s picture the mirror is the mind, and in Vallor’s the mirror is the machine.
6. Emily Bender, Timnit Gebru and others, “On the Dangers of Stochastic Parrots” (2021). LLMs stitch together text from training data without meaning. Closest technical relative of Vallor’s view. See also gebru-machine-god-doom-as-marketing in this wiki for Gebru’s critique of the “alien intelligence” story (not checked against this talk).
7. Counter-voices. Geoffrey Hinton and others argue that predicting the next word well requires modelling the world, so “mere reflection” understates it. The Othello-GPT line of work (Part 1) is the empirical edge of this dispute.

Where others push back

• Interpretability researchers: internal structure may be more than reflection (Part 1).
• Functionalists: if behaviour is indistinguishable across tasks, calling it “just reflection” is a verbal choice.
• Metaphor critics: any single metaphor hides things. A mirror does not train on feedback, take actions in tools, or run in loops (agents, 2025).

What she leaves out. Agentic systems that act in the world; multimodal and robotics models; how reinforcement learning on verifiable tasks (2024-25 reasoning models) changes training beyond imitating text. This lecture predates most of that, so the omission is partly timing (my assessment).

Check yourself

1. Name the three parts of the mirror analogy and their AI equivalents.
2. Why does she welcome the absurd errors?
3. What would a result need to show to count against “reflection”?

2. Space of reasons, the catapult and the mountain

In short. Vallor takes the space of reasons from Sellars: the shared field where people exchange thoughts about what might justify their beliefs. Thinking needs space to move in, and AI tools could help hold that space open but mostly do not. They fill it instead, and a model rarely says it doesn’t know. A robot catapulted up a mountain is not mountain climbing, and thinking takes repeated practice of scaling peaks yourself.

🎙 In the talk. Following Sellars, the space of reasons is “the field of shared human thinking where we explore and exchange thoughts about what might justify our different beliefs”. Thinking “is action like physical limbs thoughts need space to move in”. Open minds have room; “a closed mind is boxed in immobilized and helpless”.

Her claim: AI tools “can absolutely be designed to help us hold open the space of reasons but for the most part they aren’t”. They are “a mental space filling machine”. “The one intelligent thing you’ll not hear a large language model say unless you’ve specifically prompted it to do so is I don’t really know”.

Then the contrast: “unlike the high dimensional Vector space occupied by an AI model the human space of reasons is not mathematically closed but open to the Future”. The robot catapulted up a mountain is “not mountain climbing”. Asked for its reasoning, the model gives “a completely fictional madeup path of reasons”, and the two spaces just aren’t isomorphic at all. Thinking “requires repeated practice of scaling daunting Peaks yourself”. The market wants productivity with fewer steps, partly because you are “slow” and “unpredictable” and “might question the whole task”.

📚 Deeper theory

• Sellars and McDowell (Roots). Participation in the space of reasons is learned (Bildung). If the climb is skipped, the capacity is not built. This is the real force of the catapult image.
• Faithfulness of explanations. Research on chain-of-thought reports suggests the stated reasoning of a model can differ from what drove the answer (from memory; the papers I recall are 2023-25; not checked here). That supports her second claim more than her first (that the vector space is “not isomorphic” to ours), which is a statement about architecture.
• Whitehead’s rule (Roots). The distinction she draws is between automating a technique (long division) and automating the judgment about which technique and why. Whitehead’s line is usually read the first way.
• Ironies of automation (Bainbridge). The human is left to supervise the machine without the practice that would let them catch its errors. Her courtroom and journal examples fit.
• Institutions. Education, law and science are institutions that train people in the space of reasons. A tool that makes the products of those institutions cheap but not the processes may break their function.

📚 Roots: The space of reasons

The problem. What makes a belief knowledge rather than a mere state? Sellars: it can be placed in a network of justification that others can question.

1. Socrates. Questions keep an answer from closing early. Vallor jokes about philosophers being “irritating”.
2. Immanuel Kant, “What Is Enlightenment?” (1784). “Dare to think for yourself”: immaturity is letting others think for you. A strong echo of her complaint about outsourcing.
3. Wilfrid Sellars (1956). The “space of reasons” vs the “space of causes”. Knowing is a move in a social game of giving and asking for reasons.
4. John McDowell, Mind and World (1994) and Robert Brandom, Making It Explicit (1994). Develop the idea: concepts live in this space; participation needs training. McDowell stresses that we are initiated into it (his Bildung idea).
5. Jürgen Habermas. The ideal of rational public discourse. AI that fills the public sphere with plausible but unreasoned text damages it (an extension I am making, not hers).

Key move for Vallor. The space of reasons is open (new reasons can appear). A model’s vector space is closed (it holds a fixed set of weights). The two are “not isomorphic”. When a model “explains its reasoning”, it generates a plausible story, not a trace. (Research on chain-of-thought faithfulness, 2023-25, supports the cautious version of this;† here.)

📚 Roots: Offloading and deskilling

The problem. Tools that do part of our thinking can strengthen us or weaken us.

1. Plato, Phaedrus (c. 370 BCE). King Thamus warns that writing will give “the appearance of wisdom” and weaken memory. The oldest form of her worry; the standard reply is that writing also enabled philosophy.
2. Alfred North Whitehead (1911). Progress is the growth of operations we can do without thinking. Vallor’s reply: that works when the automated step is a low-level technique. It fails when the automated step is the thinking itself.
3. Lisanne Bainbridge, “Ironies of Automation” (1983). Automating tasks leaves humans with the hard residue, but without the practice needed to do it (from memory). Cited here as an extension.
4. Bernard Stiegler. Technology as pharmakon: cure and poison. See citton-attention-gesture-machine 2.2 for his proletarianisation idea (loss of theoretical knowledge). Vallor’s catapult image is a vivid version.
5. Nicholas Carr, The Shallows (2010) and The Glass Cage (2014). Argues that web use and automation reshape attention and skill.
6. Vallor, Technology and the Virtues (2016), on moral deskilling. Practices like care, honesty and courage need exercise. If a tool takes over the practice, the virtue fades.
7. The evidence. Mixed. The MIT study is small. Calculators, GPS and writing show that tools shift skills as well as erode them. Her best argument is not that all offloading is bad but that thinking skills are built by climbing, and the AI market has no incentive to pay for the climb.

Where others push back

• Scaffolding: good tutors (human or AI) can keep the learner climbing. She concedes that some people “adapt these tools for that purpose”, but doubts most users will.
• Evidence from practice: calculators and spellcheck did not end skill, and some AI tutoring trials show gains. I have not checked specific studies.
• The “average user” claim is empirical. She states it as a fact about incentives, not a measured result.

What she leaves out. The cases where a system’s answer is verifiably right and the climb is not worth making (a lookup, a translation). Her line between “technique” and “thinking” is not drawn precisely.

Check yourself

1. What is “closed” about a model’s vector space, and “open” about the space of reasons?
2. Why is the catapult not “climbing”? What is lost?
3. Which parts of Whitehead’s rule survive her objection?

3. Time: the light cone of the past and the open future

In short. Vallor says AI mirrors are confined to the light cone of the past, and the future is utterly closed to them. Asked for art styles not yet invented, models offered extremely well-established ones, and asked for stories never written, they gave stories already told and published. Turning up the temperature only draws on a wider spread of historical data, which makes it a slightly wilder guess.

🎙 In the talk. “AI mirrors are confined to the light cone of the past the future is utterly closed to them”. Asked for art styles not yet invented, models “suggest extremely wellestablished artistic approaches”; asked for stories never written, they gave stories “told and published before”. Temperature does not change this: it “just uses a wider spread of our historical data a wider statistical distribution”, so it is “a slightly Wilder guess”.

The human side: the “temporal Horizon” of “an open and indeterminate World”; Hume’s “wart”; the Assad example (“how many of you two weeks ago thought Assad would fall this week”); Ortega’s “autof fabrication”. “AI mirrors can’t make these kinds of radically new choices they are stuck pointing backward”, so they regenerate “the most common social biases and the most widespread errors”. Quoting Coeckelbergh’s idea, they are a “time machine” that cannot “take us into the future”. Used to “write the future”, they “won’t invent that future they’ll steal it from us”.

📚 Deeper theory

• Links the Roots of this idea.
• What she gets right. Statistical learners are tied to their training distribution. Forecasting rare shifts (a regime collapse) is where data-driven prediction is weakest (Taleb’s “black swans”, 2007). The Assad example is chosen for this.
• A subtle point: humans also predict from the past. Hume’s problem applies to us. The difference she claims is not better prediction, but that we can decide to depart from the pattern. That is a claim about agency, not accuracy.
• Does “new” require “outside the data”? Philosophers of creativity (Boden: combinational, exploratory, transformational creativity; see boden-not-gizmos-but-ghosts for another Boden page in this wiki) say new combinations count as creative. Her claim fits the third type: transformational creativity, which changes the space itself. Whether models ever do this is contested.
• Feedback loops. If AI-written text re-enters training data, patterns concentrate (“model collapse”, Shumailov et al. 2024)†. That is a mechanical form of her “keeps leaking” idea.
• Prediction changes the future. Predictive policing, credit scoring and recommenders do not only forecast; they shape behaviour. Her statistical ghost of the past automated into reality is a self-fulfilling loop.

📚 Roots: Time, narrative and the open future

The problem. What is it to live in time, and to have a future that is not given?

The point. Her claim that AI is “conservative” is a claim about time: a statistical model of past data has no protention of its own.

1. David Hume, Enquiry (1748). Induction cannot be justified without circularity. Vallor uses the “wart”: nothing guarantees the future resembles the past. She does not mention the ironic fact that Hume’s problem also applies to the AI’s predictions; but she means that humans act anyway.
2. Edmund Husserl (1905-28). Time-consciousness has retention (the just-past), primal impression, and protention (the about-to-be). The “horizon” is always open.
3. Paul Ricoeur, Time and Narrative (1983-85). Narrative is how human time becomes intelligible; selves are narrated (“narrative identity”). Without stories there are “only disconnected events”.
4. José Ortega y Gasset. Humans have no fixed nature; they must make themselves (“I am I and my circumstance”). Technology, for him, is the creation of a surplus of possibilities for self-making.
5. Alasdair MacIntyre, After Virtue (1981). A life is intelligible as a narrative quest. A virtue ethicist’s version of the same thought; Vallor works in that tradition.
6. Reinhart Koselleck, “space of experience / horizon of expectation” (1979). Historical time is the gap between what we have lived and what we expect. A model that only extends the experience side flattens the gap (my gloss; not in the talk).
7. Mark Coeckelbergh, Time Machines (2021). AI as a technology that reorganises time, including by feeding the past into the future. Vallor cites him by name (caption “Mark kelberg”).

📚 Roots: How the models actually work

The problem. Is “mirror” technically defensible?

1. Next-token prediction. An LLM is trained to predict the next piece of text. Pretraining on human text gives it patterns of language, facts and styles.
2. Fine-tuning and RLHF. Later stages shape the output to human preference. This is the “polish” she mentions, and adds a second source of human influence.
3. Distribution shift. Models are strongest on inputs like the training data and weaker on inputs unlike it. This underlies her “stories never written” test and robot lands 100 miles away.
4. Temperature. Samples from the learned distribution. Higher temperature flattens it. Her description is a fair account.
5. Narrow vs general models. AlphaFold-type systems are trained on curated scientific data for one task. The case for them rests on tested accuracy, not mind-likeness.
6. What would count against her. (a) Results where models solve problems with no close training match (some maths and coding evaluations; contested); (b) evidence of causally used internal world models (Part 1). Both are active research, and I have not mapped the full literature.
7. What would count for her. Hallucinated citations in courtrooms, fabricated reasoning chains, and failure outside the training distribution.

---

Where others push back

• Models can combine ideas across fields in ways few humans do, which may look like novelty. The test “list unwritten stories” is a weak one: asking for novelty in a single sample is not the same as searching for it.
• Newer systems that search and verify (for example in maths and code) can produce results not in their training data. This is after the lecture, partly; I have not mapped specific results.
• Human novelty is also mostly recombination. Her “open future” claim needs the further step that humans can choose among options that are not recombinations.

What she leaves out. Cases where closing the future is good (safety routines, aviation). She also does not separate AI that predicts events (weather) from AI that generates culture.

Check yourself

1. Why does temperature not open the future, in her account?
2. What does the Assad example show, and what does it not show?
3. Is Hume’s problem a point for her or against her?

4. Stories, boredom and the self

In short. Vallor says narrative is how selves form and how we grasp the arc of our lives and communities. AI is being sold for stories: text polishing for journalists, plot outlines for novelists, auto-diaries, lesson plans. The centre of her lecture is empty space as a portal to a future. Her examples are a mother’s guitar, a big rock and a father’s Bible, and her lesson is that a life is an action, a construction of meaning and value. An AI bot has no future to make for itself.

🎙 In the talk. Narrative “is not only the activity through which stable reflectively aware selves come into being it’s how we grasp the full Arc of our lives and our communities”. Without it “there’s no time there’s no history only disconnected events”. AI is being sold for stories: journalism turns into “a text polisher”; novelists get plot outlines; users get auto-diaries of their “family vacation”; teachers get lesson plans; the UCLA AI course; the Atlantic’s students who “can’t” read books.

The centre of the lecture: empty space as a “portal to a future”. Her own examples are the mother’s guitar, the big rock, the father’s Bible. “I choose something and then I realized that I’ve made that chosen new thing matter”. The lesson: “my life is not a passive waiting game but an action a construction of meaning and value”. “Empty space and quiet is the field where new life grows”. Modern digital life is like “a ’90s Tetris game”.

On AI: “an AI bot has no future no life to make for itself”; “for an AI miror there is no story still Untold”.

📚 Deeper theory

• Ricoeur and MacIntyre (idea 3). Narrative identity is not decoration. A self is the one who can tell what they did and why, and be held to it. If the telling is outsourced, the answerable self thins.
• Benjamin and Winnicott (Roots). Boredom as incubation. The developmental claim (a child needs to be able to be alone) is better supported in psychology than the strong claim that boredom produces innovation; the evidence for boredom aiding creativity is modest (from memory; small lab studies, mixed replication).
• “Choosing makes it matter.” This echoes Harry Frankfurt on caring and Sartre on choice creating value, though she does not cite them. It is also an existentialist claim (value is not found in advance, but made). A critic from moral realism would say some things matter regardless of our choosing.
• Why summary-of-your-life tools bother her. A diary is a practice, not only a record. A record made for you removes the interpretive work (Borgmann’s device paradigm, idea 6).
• A fair limit. Many people with disabilities, dyslexia or a second language use writing aids to enter storytelling. Her argument is strongest against replacement and weaker against assistance.

📚 Roots: Attention, boredom and unfilled space

The problem. Digital life leaves no empty moments. Is emptiness useful?

1. Blaise Pascal, Pensées (1670). “All of humanity’s problems stem from man’s inability to sit quietly in a room alone.” The old diagnosis.
2. Søren Kierkegaard. Boredom as the root of evil, and of invention (from memory).
3. Walter Benjamin, “The Storyteller” (1936). Boredom is “the dream bird that hatches the egg of experience.” Storytelling dies as boredom’s restful state disappears. Near her argument almost word for word, though she does not cite him.
4. Donald Winnicott, “The Capacity to Be Alone” (1958). The ability to be alone without anxiety is a developmental achievement. Matches the child with the guitar.
5. Herbert Simon (1971), Tim Wu (2016), James Williams (2018). The attention economy. See citton-attention-gesture-machine 2.4 for the history.
6. Hartmut Rosa, Social Acceleration (2005) and Resonance (2016). Modernity speeds up and leaves less room for being moved by the world.
7. Her Tetris image matches Rosa: each block must be slotted at once.

Where others push back

• Boredom has no automatic route to creativity; some people just scroll, and some forms of “empty time” (rumination) are harmful.
• Tools can lower barriers for people who would never otherwise tell their story.
• Students not reading long text predates AI (the Atlantic piece makes this point); blaming AI may confuse cause and effect.

What she leaves out. Collective stories and institutions (newsrooms, archives) beyond the individual child with a guitar; and how class and time poverty shape who gets empty space at all (she says “give us space” but not who is given it).

Check yourself

1. Why does she say “the story as nothing feels unbearable”?
2. What does a mirror lack that makes it unable to “tell” a story?
3. Where is her argument strongest and weakest regarding writing aids?

5. Who can live with AI: energy, Schmidt and “innovation”

In short. Vallor asks who cannot live with AI: people targeted by mindless algorithms in Gaza, political opponents identified under authoritarian governments, and people who will die of thirst when water is diverted to cool data centres. She cites the reopening of Three Mile Island, and Schmidt’s view that efficiency efforts will be swamped and that he would rather bet on AI solving the problem. Her reading is that we gambled on survival so that corporations could extract maximum profit.

🎙 In the talk. She opens by asking who cannot: people targeted in Gaza by “Mindless algorithms”, political opponents identified under authoritarian governments, and people who “will die of thirst because the water they need is diverted to cool the new data centers”. She cites Microsoft’s deal to reopen Three Mile Island and Omaha reneging on a coal pledge after deals with Google and Meta.

The Schmidt clip: efficiency efforts “will simply be swamped by the enormous needs of this new technology”, and “I’d rather bet on AI solving the problem than constraining it”. Her reading: “we gambled on their survival so that corporations today could extract maximum profit”. And she notes the grammar: “does AI need to breathe clean air”, then “a Mindless machine has no needs”. The 2022 essay: the tech ecosystem once aimed at tackling problems; now we have “crumbling schools” and “Instagram and Tik Tok”.

In Q&A: “Schmidt wasn’t entirely wrong”; if efficiency rises “usage goes up so the net gain is is not there”.

📚 Deeper theory

• Jevons and the rebound (Roots). The IEA-based coverage in Part 1 supports the direction of her claim. The size of the rebound is empirical and contested for AI specifically.
• Schmidt’s argument is a bet, not a plan. It is an instance of “technological solutionism” (Evgeny Morozov, 2013) and relies on a counterfactual: that AI’s benefit to the climate will exceed its cost. That is testable in principle (AI for grid optimisation, materials discovery) and open in fact.
• Her linguistic point. Calling a system’s resource demand its “needs” moves moral weight from people to machines. This is the same pattern as calling a corporation’s interest “the market’s demand”. It is a rhetorical analysis (Lakoff-style framing), not a technical one.
• Distribution of harm. Omaha is an environmental-justice case: costs fall on already burdened neighbourhoods, benefits go elsewhere. Crawford’s Atlas of AI (Roots) documents the same shape in mining and labour.
• Innovation vs progress. Her Q&A point needs a standard for “progress”: improved lives for many, not just new products. This is a value judgment; she says so (“what we deserve”).
• Verification. I did not check the Gaza targeting claim, the Omaha numbers or the Three Mile Island deal (the latter I recall as Sept 2024).

📚 Roots: Political economy, energy and “innovation”

The problem. Who decides what technology is for, and who pays?

1. William Stanley Jevons, The Coal Question (1865). More efficient steam engines raised coal use. The “Jevons paradox” is exactly her point that efficiency raises demand.
2. Langdon Winner, “Do Artifacts Have Politics?” (1980). Technologies embody social arrangements. Her opening claim, “technology is human all the way down”, is in this tradition.
3. David Collingridge, The Social Control of Technology (1980). Early on, control is easy but we lack information. Late, we know the harms but the technology is entrenched. Supports her “prudent and selective” case.
4. Hans Jonas, The Imperative of Responsibility (1979). Duties to future generations. This is her question about “our children’s children”.
5. Shoshana Zuboff (2019), Kate Crawford, Atlas of AI (2021), Mariana Mazzucato. Zuboff: extraction of behavioural data. Crawford: AI’s material footprint (energy, minerals, labour). Mazzucato: the state as an investor in public-value innovation, relevant to the governments question (extensions mine).
6. Regulatory arbitrage. Her point that firms play governments against each other is the standard reason environmental regulation needs coordination. The EU AI Act (2024) is a partial answer, from memory; she does not mention it.
7. Innovation vs progress. The shift from the vocabulary of “progress” to “innovation” is a long-run change in public rhetoric; I have not checked a source for it.

Where others push back

• Tech optimists: AI efficiency per task is falling fast, and AI may help decarbonise faster than it pollutes. The IEA’s own framing notes both sides.
• Regulation sceptics: states pushing hard constraints may lose investment to others, the same arbitrage she warns of.
• The 20th-century “golden age” contrast is selective. She concedes it in Q&A: “I’m not trying to tell some nostalgic story”.

What she leaves out. Numbers. The lecture gives vivid cases but no totals for AI’s share of electricity or water. It also says little on how AI used in climate science might offset costs.

Check yourself

1. What does Jevons’ paradox predict for AI efficiency gains?
2. What is the point of her “needs” remark about Schmidt?
3. Who should set the terms of AI for public good, according to her answer?

6. Q&A: selectivity, power, journalism, democracy and “existential risk”

In short. In the Q&A, Vallor calls for a wise, prudent and selective approach, with the test being the evidence that AI is the right tool for the problem. AlphaFold is a clear case, but most of the industry is pushing AI into every corner of life to earn its money back fast. She says we hear the word innovation and never progress, we deserve better, and we need better tech journalism. On democracy and marginalised communities she says she doesn’t know the answer.

🎙 In the talk.

• Which AI, without trying the rest? She says there is “a moral need for a wise prudent and selective approach”. The test: “what is the evidence that AI is the right tool to solve this problem”. AlphaFold is a clear case. But most of the industry is “trying to make their money back as fast as possible by pushing AI into any corner of your life”. Teaching is the part lecturers “actually want more time to do”; replacing it “makes no sense except as a kind of commercial strategy”.
• Governments. Platform companies are “driving the bus”. Single governments face arbitrage; so they need to share knowledge and “take commercial opportunities away from companies that show themselves to be uninterested” in responsible practice.
• Lost the plot. “You never hear the word progress anymore you just hear the word innovation”. “We deserve better”. “Why are we not in the streets screaming about that” (the information ecosystem). She asks for “the constructive kind of mad”.
• Levers. Media wants “a cozy relationship with the tech ecosystem”; we need “better Tech journalism” and the question “what kind of AI do we deserve”.
• Democracy / marginalised communities. She says “I don’t know the answer” and doubts using the same tools: “you can’t use the Master’s tools to bring down the the Master’s house”, preferring older tools unless AI’s value is shown.
• Existential risk. Energy is “the biggest existential risk in a material sense”. The “other existential risk … in the philosophical sense” is losing “our capacity to keep our futures open”. On rogue AI: “these things are mind in the way I describ they don’t have desires they don’t have intentions”; of deceptive-AI news, “they did exactly what they were told”, and “it won’t be because the AI woke up”.

📚 Deeper theory

• Collingridge and Jonas (idea 5) on prudence under uncertainty. “Prove AI is the right tool” shifts the burden of proof, close to the precautionary principle.
• Audre Lorde, “The Master’s Tools Will Never Dismantle the Master’s House” (1979) is the origin of the adage. Lorde’s point was about feminist coalition within academic structures. Applying it to AI is Vallor’s extension; whether it fits is debatable, since movements have often used the dominant media (printing, radio, social media) to organise.
• Regulatory arbitrage and Brussels/Washington dynamics (idea 5). Her call for cross-government learning matches the practice of international coordination (OECD, Council of Europe AI treaty 2024; from memory).
• Two existential risks. The “extinction” tradition (Bostrom, Ord) focuses on agentic misalignment. Vallor’s second sense draws on Ortega and Arendt (idea 3, Roots): the loss of freedom and plurality. Her first sense is about material limits. Notice she dismisses AI “waking up”, but accepts that unpredictable systems in safety-critical roles can cause disaster.
• Her reading of the deception news is an interpretation. Safety researchers would say that “did what they were told” understates the issue: goal-directed behaviour that was not intended by the instruction (instrumental convergence; Omohundro 2008, Bostrom 2014) is the very worry. Her point that humans deployed the system carelessly is compatible with theirs; her dismissal of the underlying risk is not.
• Media capture. Her claim that tech press is “tailored as PR” is broad; evidence on specific outlets is mixed†.

📚 Roots: Virtue ethics of technology

The problem. If rules cannot keep up with technology, what guides us?

1. Aristotle. Virtues are habits built by practice (Nicomachean Ethics II). Skills of judgment need exercise.
2. Alasdair MacIntyre (1981). Practices have internal goods; institutions chase external goods (money, status). The AI market’s productivity talk is “external goods” against thinking’s “internal goods” (my gloss).
3. Albert Borgmann, Technology and the Character of Contemporary Life (1984). The “device paradigm”: devices deliver commodities while hiding the practice. Close kin of “answers without the climb”.
4. Hannah Arendt, The Human Condition (1958). Natality: humans can begin something new. See citton-attention-gesture-machine 3 (idea 1). A good partner for autofabrication.
5. Vallor (2016). “Technomoral virtues” such as humility, flexibility and courage, fit a world where nobody can know the right action in advance.
6. Contrast: the longtermist and “AI safety” tradition (Nick Bostrom, Superintelligence, 2014; Toby Ord, The Precipice, 2020) treats existential risk as extinction-scale and often as agentic AI. Vallor relocates the risk. See idea 6.

Where others push back

• Safety community: “mindless” does not mean “harmless”. Optimisation without understanding can still be capable and adversarial.
• Pragmatists: demanding evidence of the “right tool” before trying slows discovery; much of what turned out useful was found by wide exploration.
• Organisers: the “master’s tools” adage is too absolute; communities have used generative tools for translation, accessibility and archiving.

What she leaves out. Concrete policy designs (audits, liability, procurement rules, public compute). Reviewers note a similar gap in the book (Part 1, Richmond summary; I have not read it).

Check yourself

1. What are the two senses of “existential risk” she gives, and which does she rank higher?
2. What is her test for whether a use of AI should be explored?
3. How would a safety researcher reply to “they did exactly what they were told”?

Q&A

(Add questions and answers here only if worth keeping.)

---

About the source and how this page was made

• Speaker: Shannon Vallor
• Event: The Turing Lectures 2024 (3rd and final lecture, “Can we live with AI?”), Alan Turing Institute with the Knowledge Quarter, King’s Place, London
• Date: 2024-12-10 (upload date; delivered a few days earlier, Assad’s fall is “this week”)
• Length: 1:33:16
• Channel: The Alan Turing Institute
• Video: The Turing Lectures: Can we live with AI?

Related page in this folder: the same “mirror” metaphor is cited as a source in bryson-relationships-with-ai-category-failure (item 9 of its reading list). No other Vallor talk page exists yet.

Speakers

Label	Who	Role	Confidence
Vallor	Shannon Vallor	Named by the host: Baillie Gifford Chair in the Ethics of Data and AI, Edinburgh Futures Institute, University of Edinburgh; directs the Centre for Technomoral Futures; co-director of the Bridging Responsible AI Divides programme. Sole lecturer, then answers questions	High (named in the talk, matches video details)
Host	”Alex Krook” (caption spelling)	Communications team, Alan Turing Institute. Introduces and chairs the Q&A	Medium (name from captions; surname spelling may differ)
Opener	”Jody” (first name only)	Knowledge Quarter. Welcome, 0:00 to 1:36	Low (first name only; surname not said)
Audience	3 in-room askers, unnamed; online questions via Slido, one signed “MJ”	Q&A	n/a

Caption warnings. Auto-captions. Corrections used on this page: “cheering / churing / touring” = Turing; “Shannon Valor” = Vallor; “aluring Institute” = Alan Turing Institute; “Wilfred sers” = Wilfrid Sellars; “Paul” = Paul Ricoeur (likely, from “narrative storytelling” context); “huso” = Husserl; “Jose Ortega iget” = José Ortega y Gasset; “ID’s … Narcissus” = Ovid’s; “Mark kelberg / cleberg” = Mark Coeckelbergh (likely; I have not checked the quoted passage in his work); “our Baba bana” = probably Abeba Birhane (low confidence; the audio is unclear and I did not verify her saying it); “noima” = Noema (magazine); “Grock” = Grok; “Balenciaga Pope” is correct; “foundational technology” probably means foundation models; “vector space … isomorphic” is the right term. “bioc sculpture” is unclear in the captions. Factual notes on the talk (all† unless a link is given):

• She says the mirror test shows that only an infant “will confuse the thing with Its Reflection”. The classic mirror test (Gallup 1970) tests whether an animal recognises itself. She is using it loosely as a figure. -: the Google “rock a day” and “glue on pizza” AI Overviews (May 2024), the Gemini historical-image overcorrection (Feb 2024) and the Grok news story about New York and the eclipse/earthquake are real episodes as far as I recall; not checked here.
• “UCLA just announced the first completely AI generated course”: from memory this concerned a course using an AI-built textbook; not verified.
• “the news report this week” about deceptive AI systems: from memory, this is about safety evaluations of frontier models in Dec 2024 (for example Apollo Research’s scheming tests on OpenAI o1); not verified. Her reading (they did what the prompt told them) is her interpretation. The evaluators framed it differently.

---

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.