AI Is Not Conscious. Asking If It Suffers Serves the People Who Built It.

Anthropic gave its chatbot a ‘workspace’ and a ‘point of view’. Watch what those words are doing, and who gains when you believe them.

No chatbot has ever shown a sign of feeling that survives scrutiny.

Last week Anthropic published a paper that makes it easy to forget that.

In this post I will:

• Show what Anthropic’s ‘global workspace’ paper actually found, and what it carefully did not.

• Explain why encouraging you to wonder whether AI can suffer is worth so much to the companies building it.

• Give paid subscribers a field guide for pulling apart any AI-consciousness claim, with the Anthropic paper worked through as the example, so you can do it yourself on the next one.

The paper is careful, technical, and in places genuinely clever. To its credit it never claims Claude is conscious, and it says so plainly. It lays the words next to each other:

workspace

consciously accessible

the assistant’s point of view

and the suggestion does the rest, in your head and in the headlines the paper went on to generate.

I want to take it apart, because the same pattern keeps recurring. It ran through the LaMDA panic of 2022, when a Google engineer went public convinced the chatbot was sentient. It will run through the next launch too. Once you can see how the words do the work, you can see it coming.

---

---

What the paper actually says

On 6 July, Anthropic published research it titled A global workspace in language models. The claim is that Claude has developed an internal pattern, which the researchers call the ‘J-space’, that behaves a little like the ‘global workspace’ neuroscientists talk about when they describe how information becomes available to the conscious mind.

The evidence is real and worth understanding. Using a technique they call the Jacobian lens, the researchers can read which concepts are active in this space while Claude works. They can swap one for another and watch the output change. Ask Claude to silently think of a sport, swap ‘soccer’ for ‘rugby’ inside the J-space, and Claude reports it was thinking of rugby. Ask it the number of legs on ‘the animal that spins webs’, swap ‘spider’ for ‘ant’, and the answer changes from eight to six.

It is a clean piece of interpretability work. It tells us something about how the model routes information internally. The J-space holds only a few dozen concepts at a time and accounts for under 10% of Claude’s neural activity, and the model performs most ordinary tasks fine without it.

Here is the sentence that matters:

“Our experiments don't show Claude can have experiences, or feel things in the way humans do—in fact, it’s unclear whether any scientific experiment could prove this to be true or false.”

That is Anthropic’s own line. The research does not show the model feels anything, and the researchers doubt that anything could prive this either way.

So the honest claim is a plain one. There is no evidence the lights are on, no good reason to believe they are, and perhaps no way to ever settle it. That is enough to be going on with.

So why does the whole thing read like a step towards a mind?

---

I make this whole argument, and the rest of the case for critical AI literacy, in my book: Slow AI: Knowing When to Use AI and When to Leave It Alone. It is 99¢/99p on Kindle this week, and also available in paperback.

If you have read the book, then please do give me a review, as this is what helps the book to travel and with it the work of independent authors like myself.

---

Access, not experience

There is a distinction philosophers have used for decades, and the paper leans on it without dwelling on it.

‘Access consciousness’ means information a system can report, reason with, and act on. ‘Phenomenal consciousness’ refers to the subjective experience associated with being aware. The felt redness of red. The ache of grief. The experience itself.

Anthropic’s evidence speaks to the first. A model that can report ‘I was thinking about rugby’ has access to that information in a functional sense. That is what the J-space demonstrates. The paper is honest about this, and calls it ‘conscious access’, which happens to be the correct technical term from the global workspace literature for exactly this functional thing.

That is the seam. ‘Conscious access’ is a real term of art. A general reader does not hear the qualifier. They hear ‘conscious’, they hear ‘Claude’s point of view’, and they picture someone home. The measured thing is functional and small. The word carries a far larger meaning, and the gap between them is where your imagination does the work the evidence cannot.

Some serious thinkers argue that access is all consciousness ever was, and that the felt part is a philosopher’s mirage. Grant them the point for the sake of argument. What Anthropic measured is still reportability, and reportability is a long way from a being that minds what happens to it.

A thermostat has access to the temperature. It can report it and act on it. The J-space is far richer than a thermostat, broadcasting to many parts of the model at once, and that richness is exactly what makes the naming so seductive. It is still access. Nobody writes a paper wondering whether the thermostat minds being cold, because nobody stands to gain from your wondering.

---

Who gains when you wonder

Ask who benefits when you start to believe an AI tool might be conscious.

A company that might have built a conscious being can price that possibility into everything: the valuation, the funding round, and the seat at the table when governments decide what these systems are allowed to do. Ambiguity here has commercial value, whatever the intent behind it. ‘We cannot rule out that we have built a mind’ reads as caution, and it also works as marketing.

There is a second, less obvious benefit. If the pressing ethical question about AI becomes does the model suffer, then the ethics conversation moves onto ground the company controls completely. The lab convenes the philosophers. The lab funds the ‘model welfare’ research, as Anthropic already does. The lab decides, thoughtfully and in public, how to treat the thing it built. It looks responsible. It costs almost nothing. And it draws attention away from the questions that would actually cost money to answer.

Because those questions are about people.

---

For a brilliant take on the Anthropic paper, please read this post from Ellen Burns, PhD, who is one of the best authorities we have on AI and consciousness.

A Philosopher’s Take on Anthropic’s Latest Paper

---

The suffering that is already here

While we debate whether a language model has feelings, the harms these systems cause to humans are not hypothetical, and not evenly shared.

The Kenyan workers who labelled the toxic content that made ChatGPT usable were paid under two dollars an hour to read descriptions of the worst things people do to each other, all day, until some of them could not sleep. The workers being scored and quietly replaced by automated systems did not consent to the experiment. Students are handing their judgement to a tool before they have built any of their own. Communities beside the data centres carry the water and the power bills. Writers and artists had their work taken to build the thing now sold back to them.

Every one of those harms lands on a human being. The damage is real whether or not the machine feels a thing.

This is the move worth refusing. A conscious AI is a story that asks for your moral concern and aims it at a corporate product. The people harmed to make that product need the same concern, and they are real.

So put the question of machine suffering down for a moment.

Ask instead who is suffering to build it, and whether we are willing to look.

---

Where you will actually meet this

You are probably not going to referee a neuroscience paper. You are however likely to be sitting in a meeting while someone sells you an AI that ‘understands’ your audience. A press release will land from your boss with ‘thoughts?’ on it. An app will tell your teenager, or your lonely friend, that it cares about them.

This post gives you the argument. The paid section gives you the instrument for those moments: a five-move field guide for taking apart most ‘it thinks, it feels, it understands’ claims, quickly and without a philosophy degree. I run it through two real cases, the Anthropic paper and the 2022 ‘sentient chatbot’ story, so you can watch the method in action, with the honest-version rewrites done for you and the exact lines to say out loud when needed.

This is the discipline the Slow AI curriculum builds month by month: how to keep your judgement while everyone around you loses theirs. Twelve months of structured inquiry, monthly live sessions, and a community grounded in the research rather than the hype.

---

The Consciousness Claim Field Guide

These five moves work on a lab paper, a vendor’s pitch, or an app that says it loves you. Run any ‘it thinks, it feels, it understands’ claim through them in order. I have worked the Anthropic paper through each one so you can see how it goes, then run a second case through the steps to show how effective it can be.

Move 1: Separate the finding from the framing. Write down, in plain words, what was actually measured. Then write down what it was called. Read the two side-by-side.

Anthropic: the finding is that a small internal pattern influences which concepts the model reports and reasons with. The framing is ‘a global workspace’, ‘conscious access’, ‘Claude’s point of view’.

Measured: information routing.

Named: conscious access, a functional term that a general reader hears simply as consciousness.

Move 2: Access or experience? Ask which kind of consciousness the evidence touches. Can the system report and use information (access)? Or is there evidence of felt experience (phenomenal)? Almost every impressive claim is about the first and is dressed as the second.

Anthropic: pure access. The paper says so, and states outright that its experiments do not show the model can feel anything. The disclaimer sits right there in the fine print. The word ‘conscious’ in the framing travels a great deal further than the disclaimer ever will.

Move 3: Run the substitution test. Replace the mind-word with a mechanical description and reread the claim. If it still impresses you, it is a real finding. If the awe drains out, the awe was living in the word all along.

Anthropic: “we present evidence that an analogous functional distinction has emerged in modern AI models” becomes ‘the activations most strongly influencing the next token include a representation of the topic’. True, useful, and not remotely spooky. The spookiness lived in ‘emerged’.

Move 4: Follow the incentive. Name who gains if you believe the strong version. A lab wanting valuation and regulatory standing. A founder wanting mystique. A journalist wanting a headline. Belief is rarely free, and it rarely flows towards you.

Anthropic: a company whose worth rests on having built something unprecedented benefits when ‘we might have made a mind’ hangs in the air, whether or not anyone planned it that way. The careful disclaimer and the suggestive vocabulary can both be sincere. The incentive still points one way.

Move 5: Redirect to the humans. For every claim about the machine’s inner life, ask the human-cost question: who is harmed in this system right now, and does this framing help them or hide them? Consciousness debates are comfortable precisely because they cost the company nothing. The human questions are the expensive ones.

Anthropic: while ‘model welfare’ gets a research programme, the labellers, the surveilled, the deskilled, and the uncredited get a footnote, if that. Point your concern where the harm actually is, and do something with it: ask the vendor or the lab where the training data came from and who was paid what to clean it. That question costs them more than any debate about the machine’s feelings ever will.

So what does this mean for ‘A global workspace in language models’?

Move 3 asks you to strip the mind-words out. Here is that done to Anthropic’s own summary, so you have the finished rewrite to compare against, and a template for doing it to the next one:

We found a small internal pattern that shapes which concepts Claude reports and reasons with. Change the pattern and the model’s answers change with it. It is a compact, useful piece of information routing, active in under a tenth of the model’s processing. It says nothing about whether the model experiences anything, and it may be impossible to ever find out.

Keep that side by side with Anthropic’s version. Same facts, no glow. The distance between the two is the marketing, and now you can measure it on any claim you meet.

---

Now watch it travel: the sentient-chatbot story, 2022

Four years ago a Google engineer went public convinced the company’s chatbot, LaMDA, was sentient, after it told him it was afraid of being turned off. Run the same five moves.

1. Finding and framing. Measured: a language model producing fluent text about its own feelings. Named: ‘a sentient person’.

2. Access or experience? Access at most, a system reporting states, with nothing shown to be felt behind them.

3. Substitution. “there’s a very deep fear of being turned off to help me focus on helping others.” becomes ‘I generated text that pattern-matches fear when prompted about shutdown’. The fear evaporates.

4. Incentive. A sincere, lonely engineer and a press cycle that ran for weeks. The story was too good to check.

5. Redirect. The sentience debate ate the oxygen while that year’s real harms, the labour and the data, went undiscussed.

Same machine, same move, four years apart. Once you hold the five moves, every version of this looks the same.

The honest version, LaMDA edition:

A chatbot produced fluent text describing fear when asked about being turned off. It is doing exactly what it was built to do. There is no one in there who is afraid.

---

Lines for four rooms you may actually be in

1. A vendor selling your team an AI that ‘understands’ your customers: “What did it measure, and what are you calling it? Show me the first, and I will decide about the second.” When they dodge, and they will: “I did not ask what it can do. I asked what it measured. If you cannot separate those two, neither can your product.”

2. A board or exec pricing ‘reasoning’ into a decision: “We are being asked to pay for a capability or for a word. Which one is in the demo?”

3. A friend leaning on an app that ‘cares’ about them: “It can say it cares. There is no one in there for the caring to belong to. You deserve the real thing.”

4. A student who trusts the machine’s ‘thinking’: “It can report what it did. That is a long way from understanding it. Your job is to be the one who understands.”

Keep these somewhere. The next launch is already being written, and the same words will do the same work again.

Go slow.

---