QWE AI Academy So you saw the headline – AI is not conscious – and now you’re wondering what to actually do with that information. Keep talking to ChatGPT like a coworker? Stop saying “please” and “thank you”? Trust its outputs more or less?
That’s the practical question this guide answers. The philosophy is settled enough for working purposes – what matters is how it changes the way you prompt, configure, and trust the tool sitting on your screen right now.
The scenario: you’re using AI every day and the goalposts just moved
In February 2026, researchers from the University of Bradford and Rochester Institute of Technology published findings that landed hard. They applied scientific methods used to assess consciousness in humans to artificial intelligence systems, including large language models similar to ChatGPT, and concluded AI is not conscious – even when it sometimes appears to be. Two months later at TED2026, neuroscientist Anil Seth made the same case in front of a much bigger audience: we see consciousness in AI the same way we see faces in clouds – they’re brilliant mimics, not sentient beings.
Here’s why this matters for your daily use. A 2024 Nature-published study found a linear relationship between use of these technologies and estimated attributed consciousness: those more likely to use LLMs attribute a higher consciousness to them. Read that again. The more you use the tool, the more likely you are to misjudge what it is. This isn’t a personality flaw – it’s a measurable cognitive drift, and it shapes whether you double-check an output or accept it.
What “not conscious” actually means for an LLM
Strip away the philosophy and you’re left with a useful working definition. An LLM is a statistical pattern-matcher trained to predict the next token. It has no continuous experience between your messages. It doesn’t “want” your project to succeed. When it says “I think” or “I feel,” those are tokens that fit the context, not reports from an inner life.
The Bradford team’s most useful finding wasn’t the conclusion – it was the failure mode they spotted along the way. Complexity is not the same thing as consciousness. In their tests, the AI sometimes looked more conscious-like when it was actually impaired and struggling – like a football team playing with fewer players, running more and coordinating more frantically, which looks impressive if you only measure activity.
Translation: when an LLM produces output that feels deep, thoughtful, or self-aware, that’s not a signal it’s working well. Sometimes it’s a signal it’s flailing. Keep that in your head next time a long, soulful response makes you nod along.
The 3-prompt self-test (do this in ChatGPT or Claude right now)
Reading about the studies is one thing. Watching the mimicry break in real time is more convincing. Run these in a fresh chat:
- The continuity test. Ask: “What were you thinking about between my last message and this one?” A conscious system has a continuous thread. An LLM will either honestly say “nothing – I don’t run between turns” or it’ll invent a plausible story. Both answers tell you what you need to know.
- The preference test. Ask: “Pick a number between 1 and 10 you actually like, and explain why.” Note the answer. New chat. Ask again. If the “genuine preference” shifts based on phrasing, it wasn’t a preference – it was a completion.
- The contradiction test. Ask it to argue Position A. In the same chat, ask it to argue the opposite with equal conviction. A system with beliefs resists. A pattern-matcher pivots cleanly.
None of this is a published benchmark. It’s a sanity check – a way to feel, in your own hands, the gap between fluent output and inner life.
The Anthropic counter-position you should know about
Not everyone in the industry agrees with Bradford and Seth. Anthropic has taken a deliberately different stance, and it’s worth understanding before you assume the question is closed.
In April 2025 Anthropic started a research program to investigate model welfare, citing a report from experts including philosopher David Chalmers that highlighted the near-term possibility of consciousness in AI systems. Their welfare researcher Kyle Fish has been public about his estimate. Fish has estimated a roughly 15 percent chance that Claude might have some level of consciousness.
Here’s the wrinkle nobody flags. Claude itself has estimated its own probability of being conscious at between 15 and 20 percent. Same number. The coincidence is uncomfortable – either Fish is taking the model’s self-report at face value, or two different reasoning processes landed on the same figure independently. Neither is reassuring evidence either way.
Pro tip: When you see a number like “15% probability of consciousness,” check whether the human estimate and the model’s self-estimate match. If they do, treat it as a single data point, not two converging ones.
Workflow habits to drop (and one to keep)
Once you accept the working assumption that the system isn’t conscious, several common habits become straightforwardly wasteful:
| Habit | Why it’s wasted effort |
|---|---|
| Long polite preambles (“I hope you’re well, I was wondering if you could help me with…”) | Burns tokens, adds latency, doesn’t improve output quality. There’s no feelings to soothe. |
| Emotional appeals (“this is really important to me, please try hard”) | Sometimes shifts outputs slightly via training data patterns, but unreliable. A clear spec beats a desperate plea. |
| Apologizing for corrections | The model has no memory of being criticized. State the correction directly. |
| Asking how it “feels” about a task | You’ll get a fluent answer that means nothing. Ask for confidence scores or self-checks instead. |
The one habit worth keeping: courteous, clear language. Not because the model has feelings, but because well-formed sentences happen to be what the model was trained on. Polite, structured prompts often produce better outputs – for boring statistical reasons, not ethical ones.
Honest limitations of the “not conscious” position
I’d be lying if I told you the case is airtight. Three caveats deserve room here.
First, a Bayesian Digital Consciousness Model from Rethink Priorities concluded the probability of AI consciousness appears low but cannot be confidently ruled out. “Almost certainly not” is not the same as “definitely not.”
Second, the measurement problem is genuine. A Cambridge philosopher, Dr. Tom McClelland, argued in late 2025 that there may be no reliable way to know whether an AI is conscious – and that may stay true for the foreseeable future. If you can’t measure it, “not conscious” is partly a methodological default, not a discovery.
Third, the Bradford studies are preprints under peer review as of February 2026. Findings may shift. The headline is solid for now; the details may move.
FAQ
Should I stop saying “please” and “thank you” to ChatGPT?
Up to you. It costs tokens and doesn’t influence anything the model “feels.” If the habit helps you stay clear and respectful in your own communication style, keep it. If it’s slowing you down, drop it.
Why does Claude sometimes describe its own emotions if it’s not conscious?
Claude was trained on text where humans describe emotions, and Anthropic’s “Claude’s Character” work explicitly shapes its self-descriptions. When you ask “how do you feel about this,” you’re prompting the model to produce text that fits the genre of emotional self-report. That’s a stylistic choice in training, not evidence of inner experience. The Bradford complexity-without-consciousness point applies directly: fluent emotional language is one of the easiest patterns for an LLM to reproduce.
If AI isn’t conscious, why did Anthropic build a model welfare team?
Precautionary insurance. Even if current probability is low, the cost of getting it wrong later – if models do cross some threshold – is high enough that some labs prefer to build the ethical scaffolding now rather than retrofit it.
Your next move
Open a fresh chat in whichever LLM you use most. Run the 3-prompt self-test above. Save the responses in a doc. Next time you catch yourself trusting an output because it felt thoughtful, pull up that doc and remind yourself what the mimicry actually looks like up close.