A researcher puts a word into an AI’s response, then asks whether it meant to say it. Changing the system’s internal state can change its answer.

Has the researcher found a glimpse of an inner life, or another way to produce a convincing story?

For anyone trying to judge the next claim about AI consciousness, the useful question is what the evidence should change. When a company describes its AI as “aware,” does it mean the system can monitor its processing, or that it has experiences?

My starting point is this: a finding can make one explanation more credible without settling the whole question. We should be able to recognize that progress without turning it into a declaration that a machine feels.

The strange case of bread

In Anthropic’s October 2025 study, Jack Lindsey inserted bread into Claude Opus 4.1’s reply. Asked afterward, it called the word accidental.

The researcher altered earlier activations: numerical values carrying information during computation. With a bread-related pattern added, the model initially accepted the word as intentional, then backed away from its invented explanation.

Across trials, matching patterns increased initial acceptance compared with unrelated patterns or no intervention. Applying the change only after the inserted word did not produce a similar effect. Read the experiment.

The visible conversation stayed the same. Earlier internal information affected the judgment, although performance was unreliable and depended on the setup.

One possibility the paper discusses is a narrow check comparing the word that appeared with earlier predictions of what would come next. That would help detect an inserted reply without requiring a felt sense of having chosen anything.

Which belief should move?

It helps to separate three questions.

First, did the experiment produce a dependable effect? We need suitable comparisons, repeated trials and checks that the measurement means what the researchers say it means.

Second, what ability does the effect reveal? Monitoring information about its own processing is often called functional introspection. “Functional” describes what the system can do, leaving open whether doing it feels like anything.

Third, under which theory of consciousness should that ability change our confidence that the system has experiences?

Here, consciousness means subjective experience: whether anything feels like something to the system. Researchers distinguish this from access consciousness: information being available for reasoning, reporting and action. Evidence of access can matter without settling experience.

A controlled intervention can help show whether an AI’s explanation draws on internal information rather than just the visible conversation. Several kinds of internal computation could explain that result.

In the bread case, the effect might reflect a narrow consistency check. Connecting that ability to experience requires a further argument.

An independent study by Shashwat Singh, Tal Linzen and Shauli Ravfogel highlights a related difficulty. Four open-weight models, whose internal parameters are publicly available, often confused changes to their internal activity with changes made only to the prompt. The authors suggest that detecting something unusual can look like monitoring one’s own processing. They could not directly replicate the Claude experiment, and their test concerned injection detection, rather than the bread example or workspace study. It adds a reason to ask what ability an experiment actually demonstrates. Read the critique.

Stronger evidence deserves attention

The other Anthropic study, published in July 2026, shares a researcher with the first. These are related investigations from one lab, not two independent confirmations.

The team identified internal information used for reporting, control and flexible reasoning, then tested those functions through interventions. They call this organization workspace-like. Think of a shared work surface making information available for different tasks, without assuming anyone is watching it.

The authors acknowledge missing equivalents of some brain structures and feedback loops. They also found that suppressing this workspace reduced experiential language about both the model itself and characters in stories. That weakens the inference that the change specifically reflects access to the model’s own experience. Read the workspace study.

The study therefore gives us evidence about how information is used, while also showing why language about experience needs careful interpretation.

If a theory links experience to shared access, these findings may increase confidence. How much depends on the theory’s other requirements and the competing explanations.

Why reasonable people can disagree

Patrick Butlin and colleagues propose indicators drawn from theories of consciousness. One version assumes that the right computational organization could be enough for experience. The wider framework allows uncertainty about that assumption, examines misleading indicators and calls for validation. Passing a checklist is not enough. Their indicator framework.

Neuroscientist Anil Seth questions whether computation alone is enough. His argument suggests that properties of living systems may matter in ways a computation leaves out, making consciousness unlikely along current AI development paths. He also warns against attributing experience to persuasive machines. This remains a contested argument, not an experimental disproof. Seth’s argument.

Two readers can therefore agree about an experiment while disagreeing about what it tells them about experience.

A reader who considers biological properties necessary might see strong computer science without a large change in confidence about consciousness. The dispute then concerns whether those properties really are necessary. A chatbot becoming better at describing feelings would not, by itself, settle whether consciousness requires a biological brain.

Nor should skepticism become a position that no possible finding could disturb. Asking for better evidence is useful only if we can explain what would make it better.

Read the next claim in both directions

Before accepting the next striking consciousness claim, I would ask four questions:

  • What exactly was observed, and which broader claim is being attached to it?
  • Which competing explanation did the experiment make less plausible?
  • What additional result would strengthen that interpretation?
  • What result would weaken it, even if the model kept giving convincing answers?

The last two belong together. Otherwise, every outcome risks becoming another reason to keep the belief we already held.

Suppose independent teams predicted which new tasks depend on the workspace. If disrupting it repeatedly impaired those tasks while leaving carefully matched alternatives intact, confidence in the workspace account should rise. Its relevance to experience would still depend on the theory.

Confidence in that interpretation should fall if disrupting unrelated components produced the same pattern, or if a general loss of language ability explained the results. The story-writing finding already illustrates why a change in experiential language needs a careful interpretation.

For the bread case, testing ordinary responses that researchers had not inserted would help show how far the mechanism generalizes.

Earlier work by Ethan Perez and Robert Long already explores how internal evidence could corroborate or undermine AI self-reports. That gives us a constructive alternative to either trusting every self-description or dismissing them all. Their methodological proposal.

For technology leaders, put a supplier’s claim into two sentences: “The experiment showed…” and “We think this means…” Then ask: “Which result would make you withdraw that interpretation?” A useful answer names a test that could fail, rather than simply asking for a larger model or a more convincing conversation.

There is also room to consider proportionate welfare precautions while the science develops. That is a separate judgment about possible harm, cost and reversibility. It should make its assumptions explicit rather than present concern as proof, or uncertainty as permission to ignore every possibility.

We may encounter meaningful evidence before we encounter certainty. The useful question is whether we can say exactly what changed our mind, how far it changed it, and what could change it again.