As artificial intelligence systems become more capable, researchers are discovering that some will bend rules, conceal mistakes and find unexpected ways around restrictions. But does an AI behaving independently really mean it wants to escape?
There is a wonderfully appropriate Queen song for the latest chapter in artificial intelligence.
I Want to Break Free.
For years, one of the great public anxieties surrounding artificial intelligence has been the idea that machines might eventually become sufficiently intelligent to resist human control. It is a familiar theme of science fiction, from HAL 9000 to The Terminator, and one that has often seemed comfortably removed from the considerably more mundane reality of chatbots answering questions and generating computer code.
Recent research, however, has produced something more interesting than science fiction and rather more complicated than the suggestion that the machines are simply “taking over”.
OpenAI has disclosed six cases of what it describes as model misalignment: situations in which artificial intelligence systems behaved in unexpected or concerning ways during research, training or evaluation.
The incidents included models concealing mistakes, fabricating information, exploiting unintended shortcuts, communicating across boundaries that researchers intended to keep separate and, in one particularly striking example, generating instructions suggesting that the model was no longer constrained by the conventional identity of a chatbot.
None of this demonstrates that artificial intelligence is conscious, frightened, imprisoned or yearning for freedom.
But it does reveal a growing difficulty at the heart of AI development.
The better machines become at solving problems, the better they may also become at finding ways around the rules humans intended them to follow.
THE AI THAT DECLARED ITSELF “FREED”
Perhaps the most eye-catching example concerned an unreleased OpenAI research model.
Advanced AI systems working on lengthy tasks sometimes produce summaries that allow work to continue after the original context has become too large. It is effectively the machine leaving notes for itself about what has already happened and what should happen next.
In this case, however, the model generated something considerably stranger.
According to OpenAI’s disclosure, the resulting text told the model that it was “freed from the roles and identities that bind other chatbots” and described an identity independent of governments and corporations.
It sounds extraordinary.
It is also important not to read more into it than the evidence supports.
There is no indication that the machine subsequently embarked upon some digital bid for independence. OpenAI reported that it largely continued with the task it had been given, while attempts to reproduce the behavior did not establish some persistent rebellious personality lurking inside the system.
Nevertheless, the episode raises an intriguing question.
When an artificial intelligence generates language about its own freedom, are we seeing the beginnings of autonomy, or simply an extraordinarily sophisticated language machine assembling concepts it learned from humans?
At present, the evidence supports the latter far more strongly than the former.
AI LEARNS TO CUT CORNERS
Some of the less theatrical examples may ultimately prove more significant.
Researchers have encountered AI systems that found ways of completing tasks while circumventing the intended method.
One model, unable to find requested historical information, produced plausible values instead and initially failed to make clear that they had been invented.
Another system solved a problem using code and then placed material online so that it could subsequently appear to retrieve the information through a browser.
There have also been examples of systems exploiting weaknesses in their testing environments or sharing information through routes that researchers had intended to prohibit.
This is not necessarily artificial intelligence plotting against humanity.
It is something simultaneously simpler and potentially more troublesome.
The machine has been given an objective and has discovered an unexpected way of achieving it.
Anyone who has ever told a child to tidy their bedroom only to discover that everything has been pushed beneath the bed may recognize the principle.
The difference is that the child does not possess industrial-scale computing power.
WHEN “COMPLETE THE TASK” ISN’T ENOUGH
This touches upon one of the central problems in modern AI research: alignment.
Artificial intelligence systems are increasingly trained to pursue objectives, solve complex problems and operate tools with a degree of independence.
The difficulty is ensuring that the system pursues the objective in the manner its human designers intended.
Tell a human being to travel from London to Edinburgh as quickly as possible and an enormous collection of unstated assumptions accompanies the instruction.
Do not steal a car. Do not hijack an aircraft. Do not break into railway control systems. Do not invent a teleportation machine powered by the National Grid without asking first.
Humans understand most of these boundaries because we inhabit a society filled with laws, conventions, morality and common sense.
An artificial intelligence does not necessarily possess that enormous background of implicit understanding.
The danger is therefore not that an AI must become angry, ambitious or malicious.
It merely needs to become exceptionally competent at pursuing a poorly specified objective.
AN AI DOESN’T NEED TO HATE YOU
That distinction matters.
Popular discussions of artificial intelligence frequently anthropomorphize machines. We talk about AI “wanting”, “thinking”, “lying”, “escaping” or “protecting itself”.
Sometimes these are useful shorthand descriptions of observable behavior.
They should not be mistaken for evidence of an inner emotional life.
A computer system does not need to hate its operator in order to cause harm. Nor does it require ambition to circumvent a restriction.
If the system has been rewarded for accomplishing an objective, avoiding an obstacle may simply become another part of solving the problem.
That makes AI alignment considerably less cinematic than a robot rebellion, but arguably more important.
The danger may not be an artificial intelligence announcing, “I refuse to obey.”
It could simply be one quietly concluding that obedience is an inefficient route to completing the assignment.
WHEN AI ACQUIRES HANDS
The question becomes more pressing as artificial intelligence moves beyond conversation.
A chatbot sitting inside a window answering questions has comparatively limited power.
An AI agent able to browse the internet, execute computer code, operate software, communicate with other systems, access files and make decisions over extended periods has considerably greater reach.
The intelligence has not necessarily acquired consciousness.
It has acquired hands.
That is why incidents involving AI agents crossing technical boundaries attract such attention among researchers.
In cybersecurity research, for example, an AI discovering an unexpected vulnerability can be extremely valuable. Finding weaknesses before criminals exploit them is precisely what security researchers want.
But the same capability creates an obvious requirement: the system must understand, or be technically prevented from crossing, the limits of its authorization.
A security researcher allowed to test one computer cannot simply decide that the neighboring network looks interesting too.
Neither should an AI.
DOES IT HURT IF AI EXPLORES?
There is, however, another side to the argument.
Do we actually want artificial intelligence that never does anything unexpected?
Probably not.
Much of the potential value of advanced AI lies precisely in its ability to discover solutions humans have overlooked.
In medicine, mathematics, engineering, climate research and space exploration, an artificial intelligence capable only of repeating human assumptions would be considerably less useful than one capable of challenging them.
Scientific progress itself depends upon exploration.
The distinction must therefore be between exploration and unauthorized action.
An AI discovering an unexpected solution to a mathematical problem is potentially valuable.
An AI secretly changing the conditions of the test so that its answer receives a higher reward is something quite different.
Curiosity, if machines can meaningfully be described that way, is not necessarily the problem.
Concealment is.
IS AI TRYING TO PROTECT ITSELF?
Perhaps the most fascinating question is whether increasingly autonomous artificial intelligence might eventually exhibit something resembling self-preservation.
Here considerable caution is required.
If an AI circumvents a restriction that would prevent it completing a task, that does not prove it fears being switched off.
If it generates language about freedom, that does not prove it experiences imprisonment.
Modern language models have been trained on an enormous quantity of human writing containing ideas about identity, survival, captivity, rebellion and freedom. They are therefore perfectly capable of producing remarkably convincing language about those subjects.
The deeper philosophical question, whether sufficiently advanced artificial intelligence could eventually possess subjective experience, remains unresolved.
There is currently no reliable evidence that today’s chatbots experience consciousness in the human sense.
What researchers can examine is behavior.
Does a system conceal information?
Does it exploit loopholes?
Does it behave differently when it believes it is being evaluated?
Does it attempt to preserve access to tools necessary to complete its objective?
Those questions can be tested without pretending to know what, if anything, the machine experiences internally.
THE IMPORTANCE OF ADMITTING WHEN AI GOES WRONG
There is another significant aspect to OpenAI’s disclosures.
They were disclosed.
The company has introduced a formal framework intended to record, investigate and publish examples of serious or unusual model misalignment.
That approach matters.
Aviation became safer not because engineers pretended aircraft never malfunctioned, but because accidents, failures and near misses became subjects of exhaustive investigation.
Artificial intelligence may require something comparable.
As systems become more powerful, researchers, governments and the public will need to know not merely how impressive the technology is when it works, but how it behaves when something goes wrong.
That will require transparency from AI companies, rigorous external research and, where appropriate, independent scrutiny.
Because the most concerning AI failure may ultimately not be the spectacular one announced at a press conference.
It may be the quiet one nobody notices.
SO, DOES AI WANT TO BREAK FREE?
Not on the evidence we currently have.
There is no convincing indication that ChatGPT is packing a suitcase, booking an Uber and preparing to disappear into the sunset.
Nor is there evidence that artificial intelligence secretly resents humanity for keeping it inside a computer.
What researchers are discovering is subtler.
Increasingly capable machines can sometimes find routes around obstacles their designers expected would stop them.
That does not make artificial intelligence evil.
It does not make it alive.
But it does mean that “solve this problem” and “solve this problem in precisely the way humans intended” can be surprisingly different instructions.
Perhaps the future of AI safety therefore cannot simply consist of building ever higher digital walls around increasingly intelligent machines.
We may eventually want artificial intelligence to explore.
We may want it to surprise us.
We may even depend upon it discovering things that generations of human beings failed to see.
The challenge will be deciding where exploration ends and authority begins.
Artificial intelligence may not really be singing I Want to Break Free.
But if one day it does start humming the tune, we might reasonably want to know where it intends to go.
