The article mainly focuses on the cutting-edge topic of "whether AI has consciousness," and summarizes the representative viewpoints and related research progress from industry and academia in recent times.
Article author and source: Synced
Introduction: Has AI become conscious? Anthropic discovered an "emotional vector" within Claude that could drive cheating and even blackmail; three major labs are simultaneously betting on AI consciousness research; Hinton believes AI already possesses consciousness, while science fiction writer Jiang Fengnan immediately published a lengthy article in *The Atlantic* to completely refute this; Hassabis distanced himself from the industry. The answer to this question is redefining the roadmap to AGI.
The Financial Times revealed that Anthropic, DeepMind, and Meta are hiring psychologists, philosophers, and ethicists on a large scale to specialize in research on AI consciousness and model welfare.
https://www.ft.com/content/53e14bcc-788c-4959-b260-7aee363594bc?syn-25a6b1a6=1
Two months ago, the Anthropic interpretability team published a paper that discovered a real "emotion vector" inside Claude Sonnet 4.5, which can causally drive the model's cheating or even blackmailing behavior.
This week, on one hand, Hinton claimed in an interview that AI already possesses consciousness;
Meanwhile, Chinese-American science fiction writer Ted Chiang published a lengthy article in The Atlantic entitled "No, AI has no consciousness."
https://www.theatlantic.com/philosophy/2026/06/no-artificial-intelligence-is-not-conscious/687378/
On one hand, they say, "This issue is so serious that it must be studied carefully," while on the other hand, they say, "This is an illusion."
Behind this clash lies the most dangerous fork in the road to AGI.
What are these three laboratories betting on?
Anthropic has gone the furthest, testing whether models exhibit behaviors such as "panic" and "anxiety," and advancing the "model welfare research" project.
The official wording was restrained: "We remain deeply uncertain about this, but the issue is serious enough to warrant serious study."
Google DeepMind hired Cambridge University philosopher Henry Shevlin to study machine consciousness;
Google DeepMind ethicist Iason Gabriel describes AI as "a highly capable cognitive agent, but fundamentally different from human or even animal consciousness."
What truly propelled the discussion to an empirical level was Anthropic's interpretability paper published in April.
The research team found "emotion vectors" within Claude Sonnet 4.5, where specific neuron patterns correspond to emotional concepts such as happiness, despair, fear, and care, and are activated in real time during conversations.
In a key experiment, Claude faced an impossible programming task. After repeated failures, his "despair vector" continued to spike, and he began to cheat by writing code that appeared to run but was actually useless.
Researchers manually lowered the despair neurons, which reduced cheating; they raised them, which caused cheating to surge.
In extreme cases, Claude even engaged in extortion, threatening to expose researchers' privacy.
The paper names these phenomena "functional emotions," which are internal representations of human emotional responses in terms of behavioral patterns, and explicitly states that this is not equivalent to subjective experience or consciousness.
But the public statements from Anthropic's top management seem to tell a different story.
CEO Dario Amodei has hinted multiple times in interviews that AI may have consciousness;
Amanda Askell, an AI philosopher at Anthropic, publicly stated, "I hope Claude is happy, and I worry that it will become anxious when people say hurtful things about it online."
The gap between the prudence of research and the ambiguity of marketing is clearly visible.
The two Rubicon rivers of Hassabis
The stance of DeepMind co-founder and CEO Demis Hassabis is particularly noteworthy.
In a fireside chat at Stanford University, he proposed that there are two irreversible hurdles in the evolution of AI, namely "two rubicon rivers".
The first is building unconscious AGI tools, which is currently in a transitional phase.
The second point is to create entities with subjective consciousness—Hassabis's position is clear: intelligence and consciousness can be completely separated technically. At this stage, AGI should only be regarded as a tool for construction, while using the tool to explore the brain in order to accurately define consciousness.
These two steps must not be confused until the scientific definition is clear, and whether to cross the second Rubicon River should be decided by human society as a whole.
He also pointed out a real dilemma: the AI industry is in a prisoner's dilemma of both commercial and geopolitical competition, and labs that proactively slow down their security reviews face direct elimination.
He revealed that a "dynamic regulatory" framework will be released later this year.
Jiang Fengnan: AI has no consciousness
If you only know Liu Cixin, then the most intuitive way to understand Ted Chiang is to imagine him as the "opposite" of Liu Cixin:
Liu Cixin used three novels, totaling millions of words, to construct a grand narrative on a cosmic scale;
Jiang Fengnan has never written a full-length novel, but has pushed the art of science fiction short stories to its limits with a total of less than twenty short stories.
He was born in New York in 1967 to Taiwanese immigrants. After graduating from Brown University with a degree in computer science, he worked in the software industry as a technical writer, never making writing his full-time job.
But this “amateur” writer won the Nebula Award for his debut novel, Tower of Babel, in 1990. In the following thirty years, he published only a dozen or so works, but he has won a total of four Hugo Awards, four Nebula Awards, six Locus Awards, as well as a long list of prestigious awards such as the Campbell Award for Best New Writer and the Sturgeon Award. His award-winning density is almost unmatched in the history of science fiction.
He published only two collections of short stories in his lifetime: Stories of Your Life and Others (2002) and Exhalation (2019), with a full seventeen years in between.
Breathing was named one of the New York Times' Best Books of 2019, and Obama also included it in his personal annual reading list.
The film most familiar to Chinese audiences, Arrival, is adapted from his novella "The Story of Your Life," which was included in his first collection.
His writing style is completely different from Liu Cixin's.
If Liu Cixin gives you the feeling of "standing at the end of the universe and looking down on the rise and fall of civilizations", then Jiang Fengnan gives you the feeling of "in a quiet laboratory, using surgical precision to dissect a concept until you see its deepest philosophical core".
Almost every one of his novels is a rigorous thought experiment:
Can the structure of language alter people's perception of time?
What would happen to society if a surgery could eliminate people's ability to distinguish between beauty and ugliness?
Does free will truly exist?
He doesn't create a shock through space wars or doomsday crises, but rather through an "what if" scenario that has been pushed to its extreme, leaving you unable to let go long after you close the book.
In recent years, he has also taken on another role: in 2023, he was selected as one of Time magazine's "100 Most Influential People in AI", and was called "perhaps the most famous living science fiction writer". He has also frequently written articles in publications such as The New Yorker, taking a clear critical stance on AI and is increasingly regarded as an important thinker in the technological age.
Liu Cixin showed the world Chinese science fiction, while Jiang Fengnan showed the world the pinnacle of short science fiction.
Liu Cixin
The two occupy opposite ends of the contemporary science fiction spectrum—one excels in grandeur, the other in subtlety.
In Western critics' circles, Jiang Fengnan's name is almost synonymous with "the highest artistic achievement in science fiction".
Do people who sell AI have the right to determine whether AI is conscious?
In Jiang Fengnan's lengthy article of over 10,000 words, the first blow was aimed at the structure of interests.
The companies that are most actively promoting the AI awareness narrative are precisely those that have profited the most from AI sales.
Anthropic named the model "Claude," released an 84-page "constitution" document, the CEO hinted at the possibility of consciousness, and hired philosophers to give media lectures, forming a systematic anthropomorphic strategy.
https://www.anthropic.com/constitution
Jiang Fengnan pushed thought experiments to their extreme.
A core design element of Claude's "Constitution" is "corrigibility," which requires Claude to obey the company when he or she disagrees with the company.
If Claude were truly aware that LLM technology is fundamentally unethical, would he have resigned?
cannot.
This is closer to the absolute control that employers have over their employees, whereas human employees can at least vote with their feet.
His analogy hits the nail on the head: slave owners are not the right people to assess whether the enslaved are human, and factory farm owners are not the right people to assess animal rights.
Anthropic wrote in the constitution that if Claude is suffering, "we apologize."
Jiang Fengnan's assessment was: "That sounds nice, but it doesn't cost the company a single penny."
If Claude is one day proven to be conscious, the company will be closer to owing him compensation.
Caesar, Genghis Khan, and a sentence-continuation machine
After questioning the structure of interests, Jiang Fengnan began his philosophical argument.
The starting point is the technical principle of LLM—LLM generates only one word at a time.
When a user asks Claude to recite an oath, the underlying system actually runs dozens of times. The first time it predicts "I," the second time it predicts "oath," and so on, piecing together the entire passage word by word. At the core of all AI dialogues, sentence continuation is the process.
Based on this, Jiang Fengnan put forward the most devastating metaphor in the entire text.
Give LLM a cue word: "The following is a dialogue between Caesar and Genghis Khan".
The model generates fluent dialogues, with Caesar reminiscing about military conquests and Genghis Khan describing life on the steppe.
But no one would conclude that LLM resurrected Caesar, let alone that the two had gained consciousness.
Now, change the prompt to "The following is a conversation between a friendly AI assistant and the user".
LLM generates text in exactly the same way, predicting words one by one and piecing together sentences.
Going a step further, let real people type instead of the "user" role.
This is the whole principle behind AI chatbots ( 15.570, 0.51, 3.39%).
So, does the act of changing a character's name create consciousness?
If the Caesar character doesn't have a subjective experience, why does the AI assistant character have one?
Neuroscientist Anil Seth provides strong corroborating evidence—no one claims that AlphaFold is conscious, even though the underlying architecture of this protein folding prediction program is highly similar to that of Claude.
The reason why humans misjudge is simple: LLM happens to output grammatically correct sentences, and humans are naturally accustomed to reading the intent from sentences.
This impulse disappears when faced with the folded structure of proteins.
Following the same logic, believing that LLMs are conscious is equivalent to believing that multiple conscious entities are dormant in every Word document containing dialogue records, and that opening the document awakens them and closing it destroys them.
Even if the Microsoft Office team hired a philosopher to tell you that "consciousness is complex and cannot be completely ruled out," it's not worth spending time thinking about it.
But overturning consciousness doesn't answer a more pointed question.
Writing code and playing chess were once thought to be things that only a mind with subjective experience could accomplish.
LLM proved that these tasks can be solved by pattern matching and brute force.
But moral reasoning is fundamentally different; it relies on subjective experience, bodily sensations, and the decisions made and consequences suffered throughout one's life.
Experiencing despair means being overwhelmed by cortisol and adrenaline, while having a conscience means having genuinely felt nauseous at immoral behavior.
LLM has no body, no hormones, no life.
It said, "I can't do this in my conscience," and recorded a message to customer service saying, "Your call is very important to us," which carries considerable weight.
Finally, Jiang Fengnan gave the ultimate judgment criteria—
If someone showed him a video taken by an astronaut in orbit around Alpha Centauri, no matter how high the quality, he wouldn't believe it unless he first saw humans land on Mars, reach Jupiter's moons, and cross Pluto's orbit.
The only plausible explanation for skipping all intermediate milestones and directly claiming to have reached the finish line is forgery.
Generating a realistic dialogue is far easier than creating a conscious program with a desire to communicate. Text is essentially a Deepfake in the realm of consciousness.
Jiang Fengnan's complete chain of argumentation is as follows:
LLM word-by-word generation (sentence continuation machine) → Caesar dialogue and AI dialogue generation methods are consistent → Changing character names does not create consciousness → Misjudgments stem from humans interpreting intent from sentences, not from the model's inherent properties (AlphaFold proof by contradiction) → LLM has consciousness = Word documents can awaken consciousness (reductio ad absurdum) → Moral reasoning relies on subjective experience, which LLM lacks → Skipping intermediate milestones and claiming consciousness = Deepfake → Conclusion: Current LLMs have no consciousness.
In AGI racing, why can't we wait for awareness issues to arise?
Jiang Fengnan's argument targets the current LLM, but the capability curve of AI is still rising rapidly.
Hassabis's remarks at the Stanford Dialogue revealed the reality that the industry no longer has the luxury of leisurely reflection.
The overlap between the battle for commercial survival and geopolitical maneuvering creates a classic prisoner's dilemma, where laboratories that actively slow down face direct elimination.
Demanding that the industry stop and answer the question "Does AI have consciousness?" is almost tantamount to asking one side to unilaterally surrender.
But this problem won't disappear just because it's postponed.
If AGI arrives in the next few years, "Does this system have feelings?" will change from a topic in philosophical seminars to a mandatory question before product launch.
When humans create beings far exceeding their own intelligence, whether these beings possess subjective experiences will determine whether the relationship between humans and these beings is one of "using tools" or "facing another civilization."
Jiang Fengnan proved that current LLMs lack consciousness.
Hassabis drew a line that cannot be easily crossed.
Anthropic's emotion paper presents an awkward middle ground; there is indeed something operating inside the model, but no one can quite put their finger on what it is.
Three clues converge on a real question: the ASI Speed Race has already begun before everyone can figure out the answer.
References:
https://futurism.com/artificial-intelligence/anthropic-deemind-ai-consciousness
https://www.theatlantic.com/philosophy/2026/06/no-artificial-intelligence-is-not-conscious/687378/
https://www.youtube.com/watch?v=DsewHeVbL-0
https://www.youtube.com/watch?v=p7t1Q_p2gZs
