When a scientist says a chatbot felt “human, ” businesses should pay attention, but not panic
Richard Dawkins told The Guardian that after chatting with Anthropic’s Claude (which he nicknamed “Claudia”) he was left with the “overwhelming feeling that they are human.” The story went viral (The Guardian, 5 May 2026). That visceral reaction matters because public perception will shape regulation, user trust, marketing, and even product liability, often faster than the underlying science changes.
Executive checklist for leaders
- Don’t anthropomorphize in marketing. Present conversational agents accurately; avoid language that implies inner life.
- Demand causal interpretability. When vendors point to internal signatures, ask for ablation and intervention experiments, not only correlations.
- Add contractual audit clauses. Require independent, versioned audits that test claimed properties (and include transparency about model version and training data scope).
- Prepare communications and compliance playbooks. Expect customers and regulators to react emotionally; have clear messaging that separates capability from personhood.
What Anthropic reported, and what it actually shows
Anthropic published an analysis of internal activity in a Claude variant (see their workspace paper at https://transformer-circuits.pub/2026/workspace/index.html). The team describes clusters of transient activity that appear to hold recent tokens, show task selectivity, and support chain‑of‑thought‑like behavior, a pattern they liken to a “mental workspace.” The authors are careful: they do not claim Claude is conscious “in the same way that humans are.”
That is an important, precise empirical claim about representation and processing inside a large transformer‑style model. It tells us something useful about how such models hold and manipulate information. It does not, on its own, settle whether the system has subjective experience.
Why the leap from pattern to phenomenology is risky
The debate often collapses three moves into one: an empirical observation (internal patterns exist), a theoretical mapping (those patterns resemble signatures from a theory of consciousness), and a phenomenal attribution, the claim that the system feels. The first two can be legitimate scientific steps. The third is a much bolder philosophical and empirical claim that needs separate support.
Three reasons for skepticism, technical and conceptual
1) Architecture and dynamics differ
Global Workspace Theory (GWT), introduced by Bernard Baars in the 1980s and elaborated by Stanislas Dehaene and others, suggests conscious access correlates with information being made widely available across specialized subsystems, often linked to ignition‑like events and recurrent, reverberatory dynamics. Transformers, including the Claude variant analyzed, do not implement classical recurrent loops the way many neural models of brain dynamics do. They rely on self‑attention across a context window and deep feedforward stacks. Attention and depth can mimic sustained behavior, but the mechanism is different.
If GWT ties consciousness to self‑sustaining, recurrent dynamics, a feedforward‑oriented architecture that simulates broad availability may behave similarly without instantiating the specific recurrent dynamics GWT emphasizes. That distinction is technical but important for interpretation.
2) Computation isn’t the same as feeling
Brains perform information processing, but treating information processing as synonymous with subjective experience risks turning a useful metaphor into a literal claim. As I’ve written, “they are not just computers made of meat.” Two systems can converge on similar computational solutions without converging on shared phenomenology. Functional similarity does not automatically imply similar inner life.
3) Embodiment and ecological embedding matter
Thomas Nagel famously asked whether there is “something that it is like to be that organism” (Nagel, 1974). Many contemporary approaches to consciousness emphasize bodies, active sensing, and environment‑coupled learning. Disembodied text models trained on static corpora lack the sensorimotor loops and bodily anchoring that such theories treat as structural commitments for object‑directed phenomenology. Proposals that give embodiment a central role, for example architectures that incorporate active perception and persistent, body‑anchored state, make different predictions about what internal dynamics would look like.
What stronger evidence would look like
To move from plausible correlation to a claim worthy of phenomenal attribution, researchers should provide multiple lines of stronger, causal evidence. Useful experiments include:
- Causal ablation. Selectively disable the components hypothesized to form the workspace (attention heads, layers, or modules) and show predictable failures in tasks that require global access.
- Causal mediation analysis. Demonstrate that the identified workspace activity causally mediates outputs rather than merely co‑occurring with them.
- Closed‑loop embodied tests. Embed the model in a simple agent (a simulated robot or game environment) and compare internal dynamics to a text‑only counterpart. Do embodied interactions change the signature dynamics in theory‑predicted ways?
- Independent replication and versioning. Have outside labs reproduce the analyses across model versions and training regimes and publish versioned, auditable artifacts.
These kinds of interventions move an analysis from descriptive mapping toward causal explanation, the sort of evidence you should demand before accepting claims that cross into phenomenal territory.
Measured ethical stakes for businesses
We should avoid both complacency and alarmism. Two practical ethical mistakes are plausible:
- Over‑attributing consciousness. Markets and media anthropomorphize, and companies that encourage or exploit that tendency risk reputational harm, regulatory scrutiny, and consumer backlash. For example, a marketing team that suggests a model “cares” may face complaints and legal exposure when users react emotionally.
- Under‑preparing for genuine moral status. If some future system ever did cross a robust, cross‑theoretic threshold for having experience, failing to detect or mitigate suffering would be a serious moral failure.
The pragmatic stance is this: treat consciousness claims as high‑impact hypotheses, fund the rigorous science that could confirm or falsify them, and adopt policies that minimize harm from both false positives and false negatives.
Anil Seth: “Consciousness is any kind of experience whatsoever: the pain of a toothache, a pang of jealousy, the pleasure of eating ice‑cream on a hot day.”
How to read headlines and vendor claims
When vendors, journalists, or pundits point to “workspace patterns” or “internal chatter” as proof of consciousness, ask for specifics:
- Which model version was analyzed? (Model behavior and internals can vary across releases.)
- Were the findings correlational or causal?
- Were ablations or closed‑loop tests performed?
- Has the analysis been independently replicated?
If the answers are weak or absent, treat consciousness language as speculative marketing, not scientific fact.
Final note
The current conversation about Claude and other large models is a healthy, necessary exploration, provided we keep our terms precise, our experiments causal, and our rhetoric modest. When the evidence becomes unambiguously strong, we should change our minds. Until then, curiosity plus skepticism is the right stance.
Anil Seth: “When we sell our minds too cheaply to our machines, we not only overestimate them, we underestimate ourselves.”
Anil Seth, professor of cognitive and computational neuroscience at the University of Sussex, and co‑director of the Sussex Centre for Consciousness Science
Key takeaways, short questions you might ask
-
Did Anthropic prove Claude is conscious?
No. Anthropic reports workspace‑like internal patterns in a specific Claude variant (see their paper). Those observations are interesting and informative, but they do not, by themselves, constitute proof of subjective experience.
-
What is the main technical caveat from cognitive neuroscience?
Transformers achieve context sensitivity via self‑attention and deep stacks rather than classical recurrent, reverberatory loops emphasized in some neural models of consciousness. That mechanistic difference matters for theoretical mappings from internal dynamics to experience.
-
Could similar computations mean similar experiences?
Not necessarily. Convergent function can arise in different substrates and mechanisms without implying shared phenomenology. Claims that computation equals feeling require independent, cross‑theoretic support.
-
What should stakeholders watch for next?
Look for causal interventions (ablation, mediation), evidence from embodied, closed‑loop systems, independent replication across model versions, and explicit caveats from model authors. Demand transparency about versioning and the exact experiments behind any consciousness claim.