Session 1: Getting Useful Answers from AI

Know What You Are Holding

Answer-shaped objects and the human judgement that makes them useful

Written by Charli-Jo Tyrer for TAVIP, July 2026

An AI assistant gives you something that looks and behaves like an answer. Your job is to decide what kind of answer it is—and what weight it can safely bear.

Introduction

From choosing an assistant to understanding its answers

AI Assistants introduced ChatGPT, Claude, Copilot, Gemini and Grok: what they are, where to find them and how to choose somewhere to begin. This follow-up asks the question that matters once an assistant has replied: what, exactly, are you holding?

An AI response has the familiar shape of an answer. It can be clear, confident, detailed and useful. But its fluency is not evidence that it knows, remembers or understands in the way a person does. It is an answer-shaped object: a newly generated estimate built from patterns in human expression.

That does not make the object worthless. It changes the judgement required of the person receiving it. A rough description may open a door that was previously closed; the same degree of uncertainty may be unacceptable when health, money, law or safety is at stake. The useful question is not simply whether AI can be wrong. It is what happens if this answer is wrong, what alternatives are available and whether you can check it.

What this session is for

By the end of this session, you should be able to:

• explain the phrase answer-shaped object in ordinary language;
• separate the confidence and polish of an answer from its reliability;
• judge an AI response in relation to the task, the stakes and the available alternatives;
• recognise when an answer is useful as a starting point and when it needs independent checking or professional expertise; and
• keep human judgement as the load-bearing part of the process.

The route through the session

The main text began life as a sequence of postcards: short sections that each turn the same object in the hand. Together they move from how the object is made to how we decide what it is worth.

• The shape of the thing: Why a fluent response resembles knowledge without guaranteeing it.
• The Eiffel Tower test: How an AI can construct a useful description without seeing or remembering.
• Useful compared with what?: Why the real alternative may be no description, no explanation or no accessible route into the subject.
• The hard cases: Why plausibility becomes dangerous when the recipient lacks the knowledge needed to test it.
• Dead reckoning: A practical way to think about estimates, drift and different kinds of water.
• The human layer: Why experience, context and judgement remain load-bearing.
• Dismissal-shaped objects: Why blanket rejection can avoid the same careful evaluation that blind acceptance avoids.
• What we actually get: A position between treating AI as truth and treating it as worthless noise.

One idea, three versions

The session text is followed by two complete alternative versions of the same argument. The basic version is written as a continuous article for general readers. The philosophy version develops the epistemic questions in greater depth. They are not three different claims. They are three routes into the same claim, written for different readers and purposes.

The postcard sequence

Read these sections in order, or pause after each one for discussion. The repeated question is simple: what can this object reasonably help me do?

Every discussion about AI seems to circle the same argument. Is it correct? Is it hallucinating? Does it actually understand what it is saying? Those are interesting questions. They are also slightly beside the point.

Because when you interact with a large language model, what you actually receive is something simpler and stranger: an answer-shaped object.

The Shape of the Thing

A large language model takes a sequence of tokens—words or fragments of words—and predicts the most statistically likely next token based on patterns learned from enormous amounts of human writing. Then it repeats that process again and again.

From the outside, this produces something familiar: you type a question and something that looks like an answer appears. But what the system actually produced is not knowledge in the human sense. It is a probabilistic artefact shaped by the patterns of human expression.

It has the structure of an answer. It behaves like an answer. Sometimes it is a very good answer. But mechanically speaking, it is something else: a statistical estimate formed from the accumulated ways humans have talked about similar things before.

The Eiffel Tower Test

Imagine asking a model: “What does it look like from the top of the Eiffel Tower?” The response will probably be vivid. It may describe the Seine winding through the city, the pale rooftops of Paris and the way Montmartre rises in the distance.

The model has never been there. But thousands of people have. They wrote travel blogs, novels, photo captions and diary entries. The model’s training data contains those descriptions. When you ask the question, it navigates that landscape of human testimony and constructs a likely description.

What comes back is not direct experience. It is an average of human description: an answer-shaped object built from the statistical distribution of how people talk about that view.

Useful Compared to What?

Most of the AI debate asks, “Is the answer true?” But another question matters just as much: “Is it useful?”

Consider a blind person asking that Eiffel Tower question. They are not comparing the model’s answer to the real view. They cannot see the view. The comparison is between the answer and nothing at all.

Measured against nothing, the value of that answer-shaped object changes dramatically. It becomes a navigational estimate: a way to participate in a conversation about something that would otherwise be inaccessible. It is not sight. But it is not nothing.

I am blind. I have spent four decades working in access and assistive technology. When I ask a model what the view looks like from the Eiffel Tower, I have absolute domain expertise on what that answer is worth to me. I know what I am holding, its limits and how to use it.

But What About the Hard Cases?

Now change the question: “My husband seems depressed. How can I help?” The model will produce an answer-shaped object. It will probably be structured, compassionate and plausible. It may suggest listening without judgement, encouraging professional help and being patient. It will sound like good advice.

But the success function here is vastly more complex than the Eiffel Tower. I may not have the expertise to detect an answer that is subtly wrong precisely because it sounds so plausible. The convincingness of the shape scales independently of the accuracy of the content. The answer sounds most authoritative exactly where you are least equipped to judge it.

Dead Reckoning

In navigation, dead reckoning estimates your position using your previous position, direction, speed and elapsed time. The estimate drifts over time. It is not ground truth, but it is still incredibly useful when the alternative is having no idea where you are.

Large language models work in a similar way. They provide dead-reckoning knowledge: estimates derived from accumulated patterns of human expression. Their usefulness depends on the waters you are in. In open ocean with nothing else to steer by, a rough position estimate is invaluable. In a narrow harbour with rocks, the same margin of error kills you.

The question is not merely “Is this better than nothing?” It is whether the person receiving it knows what kind of water they are in.

The Human Layer

The answer-shaped object is not the end of the process. It is the beginning. A human being—with experience, context and judgement—decides whether that object is useful.

The model produces the estimate. Your wetware evaluates it. The model does not know. The model does not understand. The model generates answer-shaped objects. The human decides what they are worth.

The people for whom the answer-shaped object fills the biggest gap may also be least equipped to judge when the estimate is drifting. The human layer is not optional. It is load-bearing.

Dismissal-Shaped Objects

The loudest voices in the AI debate sometimes reduce their position to one gesture: “It is all slop. Confabulation. AI bollocks.” That refuses to think about baselines.

When someone dismisses all model output as worthless, they compare it to an idealised standard—expert knowledge, direct experience or verified truth—that many people asking the question never had access to in the first place. They can produce dismissal-shaped objects: responses that have the form of a considered position without the substance of one.

If you want to argue that large language models are dangerous, argue it properly. Show where the estimate drifts, who gets hurt when the shape deceives, and where the rocks are in the harbour.

What We Actually Get

Large language models do not give us truth. They give us answer-shaped objects: navigational estimates drawn from the vast archive of human expression.

Sometimes those estimates are exactly what you need. Sometimes they are dangerous. The difference depends on context, stakes, what you know and what you do not. What we do with them is still, unmistakably, a human job.

That job starts with being honest about what we are holding: not knowledge, not slop, but something in between—an answer-shaped object.

Further reading

The two readings below restate the postcard sequence without changing its central idea. Choose the version that matches the depth and style you want—or read both and notice what changes when the audience changes.

Basic reading: An Answer-Shaped Object

A continuous version for general readers, using ordinary language and practical examples.

How the shape is made

AI assistants have learned patterns from enormous amounts of human writing. When you type a question, they use those patterns to generate a likely response, one small piece at a time. We are used to reading confidence as a sign of knowledge, but an AI can produce confidence, detail and tidy bullet points without possessing the knowledge behind them.

That does not make everything it says wrong. It can produce excellent answers. It means the polish of the answer is not proof that its contents are true.

The Eiffel Tower test

An AI has never stood on the Eiffel Tower. Yet it can construct a fresh description from travel guides, captions, stories and other accounts. What comes back is not a memory and not sight. It is an estimate made from ways human beings have described that view.

For a blind person, the practical choice may be between a reasonable description and no description at all. Against nothing, an answer-shaped object can be enormously valuable. It gives a mental picture and lets someone join the conversation. It is not sight, but it is not nothing.

Open sea or rocky harbour?

Sailors used dead reckoning to estimate position from direction, speed and travel time. The estimate could drift, but it was better than no idea where you were.

If you ask for ideas for a birthday card, a slightly wonky answer will do no harm. That is open sea. If you ask whether a strange pain means you can safely wait until Monday, the same margin of error could be dangerous. That is a rocky harbour.

The important question is not merely whether AI can get things wrong. It can. Ask: what happens if this particular answer is wrong?

When the answer sounds better than it is

For low-stakes tasks, try it, edit it and move on. For important factual claims, ask for sources and check them. For medical, legal, financial or safety decisions, use AI to help you understand the issue or prepare questions—not as the final authority.

Your judgement is not an optional extra. The AI creates the answer-shaped object. You decide what weight it can bear.

The other lazy answer

“It is all slop. AI makes things up. Never trust any of it” is equally unhelpful. Human beings also misremember, misunderstand and repeat rumours. We do not therefore decide all human conversation is worthless.

Serious criticism asks where an answer might drift, who could be hurt and what happens when someone mistakes the map for the territory.

Know what you are holding

AI assistants do not hand us truth. They give us answer-shaped objects built from patterns in the vast archive of things human beings have written. Sometimes that object is a starting point, a translation into plainer language, a description, a draft or a way into a subject that previously felt closed.

Sometimes it is confidently wrong. The useful position lies between worship and dismissal. Know what you are holding, know what kind of waters you are in, then decide what to do with it.

Charli-Jo Tyrer is a writer, conceptual artist and accessibility thinker. She has been blind for forty years and has worked in access and assistive technology for most of her adult life.

Philosophy reading: An Answer-Shaped Object

A deeper version for readers interested in knowledge, testimony, uncertainty and warranted reliance.

Discussions of generative AI repeatedly return to whether an answer is correct, whether it is hallucinating and whether the system understands. These questions can obscure a prior one: what kind of thing does a large language model give us when it responds?

It gives us an answer-shaped object. The phrase marks a difference between an utterance with the recognisable form of an answer and something that qualifies as knowledge. An answer-shaped object may be accurate, illuminating and useful; it may also be subtly or spectacularly false. Its linguistic confidence does not tell us which.

Large language models produce probabilistic artefacts assembled from patterns in human expression. Their outputs can have epistemic value without becoming knowledge, just as a map, estimate or piece of testimony can help us navigate without guaranteeing truth.

The shape of an answer

A fluent answer is not the report of a mind consulting beliefs or remembering an experience. It is generated from learned statistical relationships within vast quantities of human-produced material. The familiar surface encourages us to import assumptions from human conversation: that confidence reflects certainty, explanation reflects understanding and detailed recollection belongs to somebody who remembers. None of those inferences is secure.

The answer-shaped object inherits forms through which people express knowledge without inheriting the ordinary human relationship to that knowledge. Its plausibility and truth can vary independently. This is both its usefulness and danger.

Value under non-ideal conditions

For a blind person seeking a visual description, a migrant trying to understand an official letter or somebody meeting specialist language for the first time, the operative alternative may genuinely be nothing. A model-generated description may enable participation; it does not remove the obligation to make original information accessible or provide human expertise where the stakes require it.

The baseline must illuminate the value of access without becoming a ceiling on what people are entitled to receive.

When plausibility outruns warrant

The user can receive a polished conclusion without being able to inspect how it was produced or which sources support it. This is epistemic opacity and a problem of unequal evaluative resources. The human layer is load-bearing, but the responsibility is not evenly distributed.

Accessibility can increase epistemic agency, but access to an answer is not the same as access to the means of evaluating it.

Dismissal-shaped objects

To say that a model can fabricate does not establish that every output is worthless, any more than the fallibility of testimony establishes that nobody can tell us anything. A blanket dismissal can have the rhetorical appearance of a considered position while giving little account of context, alternatives or consequences.

The useful questions are: where does the estimate drift, who is likely to rely on it, what other source was realistically available, what happens if it is wrong, and where are the rocks in the harbour?

What we are holding

Large language models generate answer-shaped objects: linguistic artefacts that may transmit, recombine or distort elements of the human archive from which their patterns were learned. Sometimes such an object is exactly what a person needs. Sometimes its fluency disguises an unacceptable risk.

The difference depends on the question, stakes, available alternatives and the recipient’s capacity to evaluate the result. Between knowledge and worthless noise lies a large, familiar territory: estimates, maps, testimony, hypotheses, heuristics and provisional representations. Generative AI creates a new and unusually persuasive member of that category.

The interesting question is whether we understand what we are holding, what it permits us reasonably to believe and what we should do next.

Charli-Jo Tyrer is a writer, conceptual artist and accessibility thinker. Her work examines accessibility, artificial intelligence and the integrity of information systems.

About this handbook

Author and organisation

This handbook was written by Charli-Jo Tyrer for TAVIP in July 2026 as a follow-up to AI Assistants in the Confident with AI course.

Licence

This handbook is licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). You may copy, share, adapt and use it for any purpose, including commercially, provided that you credit the author.

Suggested attribution: Know What You Are Holding, written by Charli-Jo Tyrer for TAVIP, July 2026. Licensed under CC BY 4.0.