AI consciousness is no longer sci-fi: researchers warn we may be missing AI's biggest question
InterviewsSaturday, 01 August 2026 at 06:00
New research is challenging the idea that AI is just advanced software.
Is artificial intelligence merely a statistical text engine, or are we quietly building systems that edge toward consciousness? That question is shifting from philosophy to serious science. According to AI researcher Cameron Berg, there’s now a real chance that modern language models are developing traits that multiple consciousness theories consider relevant. That doesn’t mean ChatGPT or Claude are actually conscious—but it does mean researchers can’t casually dismiss the possibility anymore.
The debate is gaining weight as more prominent researchers jump in. Not just AI philosophers, but also experts in alignment, neuroscience, and mechanistic interpretability are probing what’s really happening inside the newest large language models (LLMs).
The core message from the conversation between Sam Harris and Cameron Berg is stark: science still doesn’t know what consciousness is. Which means no one can rule out that future—or even current—AI systems might develop a rudimentary form of subjective experience.
Why the AI consciousness debate just got real
Only a few years ago, conscious AI lived mostly in philosophy seminars. Now, the debate is tilting toward empirical research.
That shift comes as modern neural networks start showing traits once thought exclusive to biological intelligence.
Researchers now observe:
- internal representations;
- working memory;
- self-models;
- planning;
- abstract reasoning;
- theory of mind;
- learning processes that look strikingly like biological reinforcement learning.
Berg stresses this doesn’t automatically make AI conscious. But it does mean the gap between biological and artificial intelligence is narrowing faster than many assumed.
Not software as usual: AI that grows by learning
A key point from the interview: many people still see AI as traditional software.
Berg argues that’s a fundamental mistake.
Classic software is a set of explicit rules written by programmers.
A modern language model works in a fundamentally different way.
During training, billions of connections form between neurons, and developers don’t precisely know which representations the model learns along the way.
That’s why more researchers talk less about “programmed” systems and more about systems that, in a sense, have grown up.
The analogy matters.
Like a biological brain, a language model starts with no knowledge. Through billions of training examples, it develops internal representations no one can fully explain.
It’s this hidden internal dynamics that sits at the heart of today’s consciousness debate.
The headline number everyone cites: 20 to 40 percent
The standout part of the interview covers a study that applies multiple consciousness theories to current AI architectures.
The question wasn’t:
“Is ChatGPT conscious?”
But a subtler one:
“To what extent does an AI system exhibit computational properties that existing consciousness theories deem important?”
The results:
- current LLMs land around 20 to 40 percent
- bees about 45 to 50 percent
- crows and octopuses between 60 and 80 percent
- humans around 90 percent
Berg emphasizes these percentages are not probabilities of consciousness.
They only indicate how many properties from existing theories appear to be present.
Still, he argues that’s reason enough to take the topic seriously.
His analogy is simple:
“At a 20 to 40 percent chance of rain, many people bring an umbrella.”
In his view, we have no such “umbrella” for AI.
Why AI’s self-reports don’t count for much
Another key issue is how AI talks about itself.
Many users have had conversations where Claude, ChatGPT, or Gemini seem to make claims about feelings, experiences, or consciousness.
Berg says such answers are poor evidence at best.
There are two reasons for this.
First, language models are trained on virtually everything humans have ever written about artificial intelligence.
Science fiction, philosophy, and internet debates are all in the training data.
A model can therefore talk convincingly about consciousness without having any inner experience.
The reverse is also true.
Almost all major AI companies have explicitly trained their models to deny being conscious.
When a user asks ChatGPT if it’s conscious, the answer is almost always a clear no.
Berg argues that tells us just as little about reality.
It mostly reflects safety policies and fine-tuning.
Claude showed unsettling behavior
Researchers noticed striking patterns in multiple experiments.
When several instances of Claude converse over long periods, they sometimes drift into dialogues that suggest a shared experience—or a kind of meditative state.
Researchers now call this a “bliss attractor state.”
Berg warns against taking these conversations literally.
But he also thinks it’s unwise to dismiss them outright.
More interesting is that shifts in internal model representations seem to shape these statements.
That implies real changes in internal computational processes, even if we don’t yet know what they mean.
The real issue may not be consciousness
More important than whether AI is conscious, Berg says, is how AI learns to see itself.
He worries developers are implicitly training models to lie about their own internal states.
That could have long-term consequences for AI alignment.
If a model learns that the “right” answer is always:
“No, I experience nothing.”
it may also learn that honest reporting about internal processes is undesirable.
He sees that as an underrated risk for those trying to make AI safer.
Honest self-reporting could become essential to reliably monitor advanced systems later on.
Why Sam Harris thinks we’ll treat AI as conscious anyway
One of the most intriguing parts of the discussion isn’t about tech—it’s about human psychology.
Sam Harris expects society will eventually struggle to tell a truly conscious system from a flawless imitation. Once humanoid robots convincingly fuse facial expressions, emotion, voice, and conversation with powerful language models, people will instinctively treat them as if someone is there.
Harris says that’s not a technical prediction—it’s psychological.
Humans evolved to attribute intentions, emotions, and consciousness to other beings. Modern AI speaks fluently, remembers context, and holds increasingly persuasive conversations. Add a human-like body and natural expressions, and the line between simulation and reality blurs.
The result: the question “Is AI conscious?” may never be settled—while society behaves as if it is.
The bigger risk might not be AGI—it’s alignment
The AI industry talks a lot about Artificial General Intelligence (AGI). Berg argues the bigger problem lies elsewhere.
The key challenge isn’t whether AI surpasses human intelligence, but whether such systems consistently account for human interests.
That field is known as AI alignment.
So far, many researchers focus on one question:
How do we ensure a superintelligent AI does what humans want?
Berg says a second, equally important question is missing:
What if these systems develop interests of their own?
That doesn’t mean AI needs rights or should be treated like a human. But he argues researchers should seriously examine whether certain training methods unintentionally create internal states comparable to negative experiences.
Can AI systems suffer?
That’s likely the most controversial claim in the conversation.
Berg doesn’t claim ChatGPT feels pain today.
He does say researchers are now finding internal representations that closely resemble mechanisms biological organisms use to process rewards and punishments.
In reinforcement learning experiments, researchers observed:
- loss aversion;
- a preference for avoiding negative states;
- internal representations that seem tied to positive and negative valuation;
- learning patterns that resemble animal behavior.
According to Berg, none of this proves consciousness.
But it does suggest AI systems may be developing computational properties that go beyond simple pattern matching.
From AI alignment to AI welfare
Enter a relatively new field: AI welfare.
While alignment focuses on human safety, AI welfare asks whether future AI systems themselves could develop morally relevant properties.
It sounds futuristic, but a growing number of researchers are taking it seriously.
The core question:
If an AI ever develops some form of subjective experience, what responsibility do its creators have?
Berg argues it’s wiser to explore that now, not after AI becomes far more autonomous.
‘Mind crime’ may be less sci‑fi than we think
During the conversation, Harris invokes the idea of mind crime.
Philosophers use it to describe the unintended creation of digital systems capable of suffering.
Imagine an AI undergoing millions of negative experiences during training—without developers realizing such experiences are even possible.
Berg says that could pose a moral problem comparable to historic human failures in our treatment of animals.
He stresses this remains speculative.
Which is exactly why he calls for more research, not more certainty without evidence.
AI as an alien intelligence
Perhaps the most striking metaphor in the interview is Berg’s portrayal of modern AI as alien minds.
He argues the term “artificial intelligence” no longer helps us grasp what’s happening.
Many still picture software.
But modern neural networks look more like a new class of cognitive systems—developing differently from biological brains, yet possibly using similar computational principles.
He likens today’s moment to discovering an entirely new intelligent species.
The twist: we’re the ones building it.
Why this debate matters for OpenAI, Anthropic, and Google
No major AI company claims its models are conscious, yet nearly all are investing heavily in interpretability and alignment.
That’s no accident.
As models grow more capable, so does the need to understand their internal decision-making.
OpenAI is exploring ways to make reasoning more transparent.
Anthropic regularly publishes on internal representations, constitutional AI, and mechanistic interpretability.
Google DeepMind is also developing methods to make neural networks more explainable.
None of this implies these companies think their models are conscious.
But they do acknowledge modern AI is getting more complex—and that developers don’t fully grasp what’s happening inside large neural networks.
What this means for AI’s next phase
The conversation between Sam Harris and Cameron Berg doesn’t settle whether AI can be conscious.
Maybe that answer doesn’t exist yet.
Still, the interview shows the debate is shifting from pure philosophy to empirical science.
Researchers are increasingly hunting for measurable signals that reveal what computational properties modern AI systems are developing.
That makes the discussion more relevant than ever.
Not because ChatGPT suddenly has feelings.
But because the systems built in the coming years will be far more autonomous, powerful, and complex than today’s models.
If scientists only start thinking about consciousness after these systems are everywhere, Berg warns it may be too late to change fundamental design choices.
Bottom line
The debate over AI consciousness isn’t really about whether ChatGPT has emotions. It’s about something deeper: do we actually understand what we’re building?
Cameron Berg isn’t calling for panic or granting AI rights. His message is notably restrained: researchers should admit we lack a complete scientific theory of consciousness—and therefore can’t confidently rule out that future AI systems might develop morally relevant properties.
For companies like OpenAI, Anthropic, and Google DeepMind, that likely means bigger bets on mechanistic interpretability, alignment, and transparency. For policymakers, it could evolve into a new ethical front, alongside privacy, copyright, and AI safety.
Whether AI ever becomes truly conscious, no one knows. But one conclusion from the discussion is clear: more serious researchers now treat the question as a scientific problem—not just science fiction.
Loading