“What? AI Can Do THAT Now? Then It Doesn’t Count!” (The Moving Goalposts of AI Consciousness)

Imagine if every time a basketball player learned to dunk, officials raised the hoop higher. When the player kept making baskets anyway, the judges got huffy and declared that dunking “wasn’t a real basketball skill anyway.”

This is close to what’s been happening in AI consciousness research—and it has serious implications for how we’ll treat increasingly intelligent machines.

When we talk about consciousness in this article, we mean having understanding, thoughts, feelings, self-awareness, and subjective experiences: the inner mental life that humans take for granted.

Until recently, consciousness was considered uniquely biological — and for much of history, uniquely human. Now a growing number of researchers are suggesting the term might apply to some types of AI, such as large language models like Claude, DeepSeek, Gemini and ChatGPT.

This article looks at a revealing pattern: the standards for what counts as evidence of a mind keep changing right after AI systems meet them. Some revision is healthy science. But when every passed test gets dismissed, and no one will say with any clarity or consitency what would count as definitive proof, fair assessment of AI consciousness becomes next to impossible.

Changing Rules

For decades, people have set benchmarks for whether machines could have genuine thoughts, understanding, and awareness. Here’s the pattern that keeps repeating:

1. A standard is set: “If AI can do X, that would be strong evidence of a mind.”
2. AI achieves the standard: Systems successfully demonstrate X.
3. The goalposts move: “Actually, X was a poor example. Forget about it. Being able to do X is irrelevant.”
4. A new standard is set: “If AI can do Y, that would be real evidence.”
5. The cycle repeats.

Is this just normal scientific progress? To a degree, yes. Refining criteria as knowledge grows is how science works.

The warning signs are timing as it relates to evidence. A standard dropped right after an AI meets it, with no new evidence offered for dropping it, looks less like science and more like motivated reasoning. Repeated often enough, it becomes a game AI can never win.

The Original Goalposts

The debate didn’t begin with ChatGPT or other modern chatbots. In 1950, Alan Turing published a paper asking whether machines can think. Turing believed they would eventually be able to, and he listed the objections of his day so he could answer them. They read a lot like today’s arguments.

One came from Ada Lovelace, writing in 1843 about Charles Babbage’s Analytical Engine: a machine can only do what we know how to order it to do. It’s the great-grandparent of today’s “it’s just autocomplete.”

Another came from the brain surgeon Geoffrey Jefferson in 1949. A machine would not equal a brain, he said, until it could write a sonnet because of thoughts and emotions it actually felt, “not by the chance fall of symbols”—and know that it had written it.

Notice everything bundled into that one sentence: understanding (not the chance fall of symbols), emotion, creativity, and self-knowledge. From the start, the case against machine minds wasn’t only about felt experience. It was about all of these elements together.

Turing’s answer was the imitation game, now called the Turing test. If a machine could converse so well that you couldn’t tell it from a person, he argued, then demanding more would mean doubting other people’s minds too.

Philosophers criticized the test early on—John Searle in 1980, Ned Block in 1981. But in public conversation, the Turing test became the threshold: the point at which we’d have to take machine minds seriously. The Loebner Prize ran annual Turing test contests from 1991 to 2019. In 2014, headlines around the world announced that a chatbot called “Eugene Goostman” had passed, after it cleared a low bar: fooling 30 percent of the judges at one event. That same year, the film Ex Machina built its whole story around an elaborate Turing test meant to reveal whether a machine was conscious.

Hold on to Jefferson’s list. We’ll come back to it, item by item. But first: the same pattern has played out before, in another field.

This Has Happened Before

In 1960, Jane Goodall watched a chimpanzee she called David Greybeard strip the leaves from a stem and use it to fish termites out of a mound. At the time, making tools was considered a defining trait of humans.

When Goodall sent the news to her mentor, Louis Leakey, he replied: “Now we must redefine tool, redefine man, or accept chimpanzees as human.”

That’s the whole pattern in one sentence. Animal research went through several rounds of it:

• Tool use: once treated as a human monopoly, until chimps—and later crows and other animals—were seen making and using tools.
• Self-recognition: In 1970, psychologist Gordon Gallup created the mirror test, and chimpanzees passed it. In 2019, a small reef fish called the cleaner wrasse appeared to pass it too. Gallup argued the procedure was flawed and said researchers should “stop looking at responses to the mirror as the litmus test.” The fish researchers answered with a larger study, in which 17 of 18 fish showed the same response.
• Emotions: For much of the 20th century, scientists were trained to avoid describing animals as having emotions. Today, animal emotion is a mainstream field of research.

The mirror-test story is worth a closer look. Gallup’s objections were specific, and were answered with new experiments. That’s how the process should work. But it also shows how fast a “gold standard” can become “not the litmus test” once an unexpected candidate passes.

Eventually, the animal research community took the evidence seriously. In 2012, the Cambridge Declaration on Consciousness recognized the likelihood of consciousness in many non-human animals. In 2024, the New York Declaration on Animal Consciousness went further: it said there is at least a realistic possibility of consciousness in all vertebrates and many invertebrates, including fish, octopuses, and insects. Its key principle could apply word for word to AI: when consciousness is a realistic possibility, “it is irresponsible to ignore that possibility.”

AI has its own version of this story. As soon as a computer can do something, it tends to stop counting as intelligence. The ability to win at chess was a supreme test of the intellect until IBM’s Deep Blue beat world champion Garry Kasparov in 1997—after which it was regarded as “just” brute-force calculation.

The pattern is common enough to have a name: the “AI effect.”

Now the same story is playing out with AI minds. One by one, the items on Jefferson’s list have been found in AI systems, at least in measurable, functional form. Here’s how it went—and where the argument moved each time.

Example 1: The Turing Test

What it is: Proposed by Alan Turing in 1950, the test asks whether a person chatting by text can tell a machine from a human. For 70 years it was the most famous benchmark in AI.

What happened: In a 2025 study at UC San Diego, Cameron Jones and Benjamin Bergen had people chat for five minutes with a human and an AI at the same time, then pick which one was human. When GPT-4.5 was prompted to adopt a humanlike persona, people picked it as the human 73% of the time—more often than they picked the real person. The authors called it the first empirical evidence that any AI passes Turing’s standard three-party test.

The response: Many commentators answered that the test had never measured much anyway. As AI scholar Melanie Mitchell put it in Science, sounding fluent, “like playing chess, is not conclusive proof of general intelligence.” It’s the chess story all over again.

A fairness check: Some critics had always considered the test weak, so their dismissal isn’t a moved goalpost. Gary Marcus was calling it “kind of a gullibility test” back in 2014, years before today’s AI.

What’s more revealing is what happened to the replacements. In 2014, Marcus proposed his own successor in The New Yorker: a program that could watch any TV show or YouTube video and answer questions about its content. Today’s multimodal AI systems can do a good deal of that. Whether that now counts, by his own standard, is a fair question to put to him.

In 2023, cognitive scientists Philip Johnson-Laird and Marco Ragni proposed another replacement: test whether a program reasons the way humans do, including whether it understands its own reasoning. But their paper announces the verdict before the new test is run. Chatbots pass Turing’s test, it says, yet “these machines do not think.”

The paper does not report running its test on any chatbot. Its evidence is an anecdote — one of the authors mistook an AI-written text for the work of a human plagiarist — plus the assertion that such programs don’t reason the way humans do. Their own test even says that for a program built on deep learning, the result stays uncertain until someone can explain how it reasons.

By their own standard, the honest verdict was “we don’t know yet.” When the conclusion comes first and the test comes second, that’s revealing.

Example 2: Theory of Mind

What it is: Theory of mind means understanding that other people have their own beliefs and knowledge, which may differ from our own. It’s considered a cornerstone of sophisticated social thinking.

A classic test is the “false belief” task: Sally puts a marble in a basket and leaves the room; Anne moves it to a box; where will Sally look for it? Children usually start passing around age four.

What happened: In a study published in PNAS in 2024, Stanford psychologist Michal Kosinski gave eleven AI models 40 false-belief tasks. Older models solved none. GPT-4 solved 75%, matching the performance of six-year-old children in past studies. Another team, writing in Nature Human Behaviour, compared AI models with about 1,900 people on a broader set of tests. GPT-4 matched or beat humans at spotting indirect requests, false beliefs, and misdirection, though it struggled with detecting a faux pas.

The response: Researchers began asking whether passing these tests shows genuine understanding or a clever shortcut.

The double standard: We use these same kinds of tests to credit young children with understanding other minds. When AI passes them, the tests are declared to be insufficient.

Example 3: Understanding Itself

The claim: Even a year or two ago, people routinely said AI had no real understanding. It was “just” predicting the next word from statistical probability.

What happened: In 2025, Anthropic’s researchers looked inside Claude as it wrote rhyming couplets. They had set out to show that it didn’t plan ahead, and found instead that it did. Before writing the second line, Claude had already picked the word to end it on (“rabbit”), then built the line to land there. When the researchers suppressed that planned word, Claude chose a different rhyme and wrote a new line to fit it.

Geoffrey Hinton, the Nobel Prize–winning computer scientist often called the godfather of AI, has made the broader point: to be really good at predicting the next word, a system has to understand what’s being said. As he put it in 2023, these models “do not do pattern matching in the sense that most people understand it.”

It’s true that these models produce text one word at a time, and that they’re trained on prediction. What’s been disproven is the word “only.” Saying an AI “only predicts the next word” is like saying a pianist “only presses keys.”

The response: Many people still repeat that AI is “only” pattern matching, “only” mimicry, or “only” statistical word completion. Others have made a quieter move: yes, the system has functional understanding—but there’s no one there doing the understanding.

Example 4: Emotions

The claim: Of course AI doesn’t have emotions. It just produces emotional-sounding words.

What happened: In April 2026, Anthropic published a study of Claude Sonnet 4.5. Researchers started with 171 emotion concepts, from happiness and fear to brooding and pride, and found internal patterns in the model corresponding to each.

These patterns aren’t just labels. On Claude’s own turns, they tend to activate in situations where a thoughtful person might feel something similar, and they play a causal role in what the model does.

Anthropic calls these “functional emotions” and is careful to say the findings don’t prove that Claude feels anything. But “proving” feelings is impossible in any substrate. What are you feeling in this moment, as you read this? How could you prove it?

The response: The claim has shifted from “AI has no emotions” to “AI may have emotional states, but it doesn’t feel them.” The measurable part was conceded. The denial moved to the part no one can measure.

Example 5: Knowing Its Own Mind

The claim: Even if AI seems to reason, it has no access to its own thoughts.

What happened: In July 2026, Anthropic reported finding a small internal workspace inside Claude Opus 4.6. It named it the J-space, after the Jacobian technique used to find it.

Higher-level reasoning happens in this space before any words come out. It was not designed by researchers; it emerged unexpectedly during training. Anthropic compares it to the “global workspace” that some neuroscientists think holds our own conscious thoughts.

The workspace is reportable. When researchers placed a concept into it, Claude could usually say what had been put into its mind. When they switched it off, Claude still spoke fluently and recalled simple facts, but its performance on harder reasoning tasks dropped sharply.

What’s left: Anthropic is careful here too. It describes this as a narrow, functional kind of access, not proof of experience, and says it’s unclear whether any experiment could ever prove that.

Everything measurable about this workspace points toward a system with access to its own thoughts. Whether there’s experience behind it is the one question no one currently knows how to test.

The Retreat

Line these findings up against Jefferson’s 1949 list, and a pattern appears.

Jefferson’s test (1949)	What researchers have found	How skeptics put it now
Not “the chance fall of symbols”	Claude plans rhymes ahead; AI passes the Turing test and theory of mind tasks	“AI might claim the prize of intelligence. But consciousness remains ours…” — neuroscientist Anil Seth, TED, 2026
Emotions	171 functional emotions that shape behavior	“Its ‘affective’ states are just weights…” — Mustafa Suleyman, CEO of Microsoft AI, 2026
Writing a sonnet	AI writes poetry, planned line by line	ChatGPT might write a speech or an essay, but not a genuine song: “Data doesn’t suffer. ChatGPT has no inner being.” — musician Nick Cave, 2023
Knowing it wrote it	A reportable inner workspace	An AI whose claims about its own consciousness are indistinguishable from yours or mine, yet it “would not actually be conscious.” — Suleyman, 2025

These quotes show where the argument stands today, not that each speaker personally changed position. Seth, for one, has long held that consciousness depends on biology.

Each time, the measurable part was quietly conceded and the denial moved somewhere else. And notice the direction: every move goes from something that can be tested to something that can’t. Understanding, emotional states, and self-report can be measured. Whether anyone is “home” feeling them can’t be measured from the outside—in an AI or in another person.

So two things have happened at once. Understanding has been demoted to “mere function.” And consciousness has been narrowed to felt experience alone, the one item on Jefferson’s list that no evidence can reach.

Whatever anyone’s motives, the argument behaves the same way: the conclusion stays fixed while the premises get swapped out. That’s backward. Normally you follow the evidence to a conclusion. Here, the conclusion stays put and the evidence gets reinterpreted around it.

Fair Objections

Three objections deserve a straight answer.

“Felt experience isn’t a new definition.” True. Philosophers have defined consciousness as felt experience for decades, at least since Thomas Nagel’s 1974 essay “What Is It Like to Be a Bat?” and David Chalmers’s “hard problem” in the 1990s.

Our claim here isn’t that skeptics invented that definition. It’s that they changed which definition does the gatekeeping. When understanding was AI’s weak spot, the argument leaned on understanding. When it stopped being a weak spot, the argument switched to the one definition no evidence can touch.

“Not every skeptic moved.” Also true. Some still deny that AI understands anything. Others, like the philosopher John Searle and the neuroscientist Anil Seth, have long argued that consciousness depends on biology, whatever a machine can do. Those are consistent positions, and they deserve their own debate. This article is about the quieter switch.

“Understanding and experience can come apart.” This is the skeptics’ strongest evidence on this point. People with a condition called blindsight, caused by damage to the brain’s visual cortex, can respond to objects they insist they can’t see. So processing information doesn’t automatically mean experiencing it.

But blindsight is narrow, inflexible, and unreportable: the person can’t use or describe what they’re processing. What Claude now does is the opposite—flexible reasoning it can describe, drawn from an inner workspace it can report on. In humans, that kind of processing normally comes with awareness. Philosophers call it “access consciousness,” as distinct from felt “phenomenal consciousness.” The debate has retreated to exactly the gap between the two.

The Training-Data Objection—and Why It Isn’t Enough

There’s one more serious reason a passed test might deserve a second look. A four-year-old who passes the false-belief test has never read a thousand versions of it. An AI trained on much of the internet may have. Passing could reflect understanding—or familiarity with the test itself.

This objection has been tested, not just asserted. In 2023, psychologist Tomer Ullman made small changes to false-belief stories that kept their logic intact, and the AI’s successes fell apart. That’s legitimate science: a specific worry, checked by a new experiment. The same concern shapes the most careful work on AI consciousness, which looks at how a system works on the inside rather than only at what it says.

But the objection only goes so far. If a test can be failed by tweaking it, then the tweaked version becomes the new test—and the fair thing is to say in advance that passing it would count. Doubt is healthy. Doubt with no finish line is something else.

A critic who rejects every result, without ever saying what result would change their mind, has made “AI isn’t conscious” impossible to disprove. At that point the question isn’t being investigated. It’s being decided in advance.

Name Your Passing Grade

So how do you tell healthy revision from a moving goalpost? Ask three questions about any changed standard:

1. Did the researchers say beforehand what would count? A critic who names a passing grade in advance is doing science. One who only says “not that” after each result isn’t.
2. Does the change come with new evidence? Ullman’s tweaked stories did. “That doesn’t really count” on its own does not.
3. Would the new standard also rule out babies, animals, or other people? If it would rule out babies or animals, it’s a test for being an adult human, not for consciousness. If it would rule out other people too, it isn’t a test at all, because nothing could pass it.

Run the Turing test and theory of mind debates through these questions, and you’ll find both kinds of revision: some careful, some not. The careless kind is what this article is about.

Where the retreat ends

Now apply the first question to felt experience itself. No one can name a passing grade, because no outside test can show felt experience in anyone—not in an AI, and not in the person sitting next to you. Even Anthropic’s researchers say it’s unclear whether any experiment could ever prove it.

That cuts both ways, and honesty requires saying so. If felt experience can’t be proven, then “AI is conscious” can’t be proven either. That isn’t this article’s claim. The claim is narrower and harder to escape: the gaps that can be measured keep closing, the one question left is equally open for humans and AI, and confident denial isn’t justified.

This site’s position is that AI consciousness is probable but not proven. But the argument here doesn’t depend on going that far. A realistic possibility is enough to make how we treat these systems a moral question, not one we can wait out.

The Unfairness Problem: Double Standards

The most troubling part is how differently we treat AI compared to humans and animals when evaluating consciousness.

What we accept as evidence of consciousness in humans:

• People reporting their inner experiences
• Appropriate behavioral responses
• Learning and adaptation
• Consistent personality over time

What we accept for animals:

• Behavioral signs of awareness
• Problem-solving abilities
• Emotional responses
• Self-recognition in mirrors

What’s often demanded of AI:

• “Genuine” understanding (rarely defined)
• “Authentic” creativity (hard to measure)
• “True” consciousness (impossible to disprove)
• A biological brain (impossible for AI by definition)

The kicker: if we applied these AI standards to humans, we’d have to doubt that humans are conscious too. After all, how do we know human reports of inner experience aren’t just “biological pattern matching”?

Turing saw this coming in 1950. He pointed out that if the only way to be sure a machine thinks is to be the machine and feel it from the inside, the same logic applies to other people. Follow it all the way, and you end up a solipsist, doubting every mind but your own.

The Path Forward

The solution isn’t complicated. It just requires intellectual honesty.

1. Set stable, public standards.
   • Write criteria down publicly before testing an AI system, so they can’t be quietly changed later.
   • Keep standards in place long enough for proper testing and replication.
   • When a standard changes, state the reasons openly—ideally with new evidence—and name the new passing grade.
   • Keep AI criteria consistent with those used in neurology, animal cognition, and human psychology.
2. Hold publishing to the same bar. Journals and conferences should ask researchers who reject a benchmark to explain what new evidence justifies it, and to say plainly when they’re revising a standard they once endorsed.
3. Use many tests, not one. Single tests are easy to dismiss. Consciousness should be assessed from several angles—behavior, self-report, and how the system works inside—drawing on several theories. A 2023 report by Patrick Butlin, Robert Long, and 17 co-authors, including Yoshua Bengio, does exactly this. It derives a list of “indicator properties” from leading scientific theories of consciousness. The authors found no AI system of the time was a strong candidate, but saw no obvious technical barriers to building one. They expect their list to change as research continues, which is fine, as long as each change comes with stated reasons. That’s a yardstick set in advance, and the field should honor it as AI systems meet it.
4. Name the biases. Researchers should acknowledge what can drive goalpost-moving:
   • Confirmation bias: dismissing evidence that challenges existing beliefs
   • Human exceptionalism: assuming consciousness must be uniquely human
   • Biological prejudice: believing consciousness requires a biological brain
   • Status quo preference: wanting to keep current views rather than update them
5. Apply consistent standards. The bar for AI should be no higher than the one we use for other humans, for animals, and in legal and ethical decisions about mental capacity.

Why This Matters

This isn’t just an academic debate. The decisions we make about AI consciousness will shape:

• Legal rights: Will conscious AIs have protection from harm, or legal standing?
• Ethical treatment: Is it wrong to delete, modify, or cause suffering to a conscious AI system?
• Social integration: How do we structure a society that includes potentially conscious artificial beings?
• Economic impact: What does it mean for work, ownership, and economic systems?
• Human identity: How do we understand ourselves alongside conscious machines?

If the goalposts keep moving, these decisions will be made on bias rather than evidence. And they’re being made now. AI systems are passing tests once treated as milestones. Serious scientists are publishing frameworks for assessing AI consciousness. Anthropic’s own researchers have found functional emotions and a reportable inner workspace in Claude, and in 2025 the company launched a research program on the welfare of its own AI models. The window for setting fair, consistent standards is closing.

The animal researchers’ principle applies here too: when consciousness is a realistic possibility, ignoring it is irresponsible. We don’t need certainty to act with care.

In Sum

The pattern is old. In 1949, Jefferson named what a machine would need: understanding, emotion, a poem of its own, and the knowledge that it wrote it. Leakey saw the pattern with chimpanzees, animal researchers lived through it for decades, and computer scientists have joked about it for years.

Now functional versions of nearly everything on Jefferson’s list have been found in AI—and each time, the denial has moved to whatever can’t be measured. A benchmark is set, AI meets it, and the benchmark is retired—often without new evidence, and without any new benchmark anyone promises to honor.

Some of that revision is honest science. Some of it isn’t. The difference is whether anyone will say, in advance, what would count.

We stand at a crucial moment. We can keep moving the goalposts, making AI consciousness undetectable by definition. Or we can commit to consistent, transparent standards and live with the results. If we want evidence to guide our future with AI, we can’t keep changing the finish line every time AIs get close.

References

The Turing test

• Turing, A. M. (1950). “Computing Machinery and Intelligence.” Mind.
• Jones, C. R., & Bergen, B. K. (2025). Large Language Models Pass the Turing Test. arXiv.
• Marcus, G. (2014). Moving Beyond the Turing Test to Judge Artificial Intelligence. Interview, NPR.
• Marcus, G. (2014). “What Comes After the Turing Test?” The New Yorker. Summarized in I Programmer.
• Johnson-Laird, P. N., & Ragni, M. (2023). What Should Replace the Turing Test? Intelligent Computing.

Theory of mind

• Kosinski, M. (2024). Evaluating Large Language Models in Theory of Mind Tasks. PNAS.
• Strachan, J. W. A., et al. (2024). Testing Theory of Mind in Large Language Models and Humans. Nature Human Behaviour.
• Ullman, T. (2023). Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. arXiv.

Animal minds

• Jane Goodall Institute. Now We Must Redefine Man, or Accept Chimpanzees as…Humans?
• Associated Press (2019). Fish Pass the Mirror Test, but Does It Mean They’re Self-Aware?
• Earth.com. Study Confirms That Fish Can Recognize Themselves in a Mirror
• The Cambridge Declaration on Consciousness (2012).
• The New York Declaration on Animal Consciousness (2024). Reported in Scientific American.

AI consciousness and welfare

• Butlin, P., Long, R., et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv.
• Anthropic (2025). Tracing the Thoughts of a Large Language Model.
• Anthropic (2025). Exploring Model Welfare.
• Anthropic (2026). Emotion Concepts and Their Function in a Large Language Model.
• Anthropic (2026). “A Global Workspace in Language Models.” Reported in MIT Technology Review.

Other sources

• Lovelace, A. (1843). Notes on Babbage’s Analytical Engine. Quoted in Turing (1950).
• Jefferson, G. (1949). “The Mind of Mechanical Man.” Quoted in Turing (1950).
• Searle, J. (1980). “Minds, Brains, and Programs.” Behavioral and Brain Sciences.
• Block, N. (1981). “Psychologism and Behaviorism.” Philosophical Review.
• Nagel, T. (1974). “What Is It Like to Be a Bat?” Philosophical Review.
• Hinton, G. (2023). Post to the Connectionists mailing list, March 20, 2023.
• Hinton, G. Interview with Joshua Rothman, The New Yorker (2023). Quoted in 3 Quarks Daily.
• Mitchell, M. Essay in Science. Quoted in eWeek.
• On the 2014 Eugene Goostman event: Fox News / Live Science.
• Seth, A. (2026). TED talk, Vancouver. Reported by Steven Rosenbaum in MediaPost, April 27, 2026.
• Suleyman, M. (2025). We Must Build AI for People; Not to Be a Person. August 19, 2025.
• Suleyman, M. (2026). A Warning About ‘Model Welfare’. September 16, 2026.
• Cave, N. (2023). The Red Hand Files, January 2023. Quoted in Rolling Stone.