What does it mean to be human? We ponder this question in philosophy seminars, in deep existential crises, and — increasingly — in the rapidly changing world of modern AI.
The question is inherently interdisciplinary: bringing together philosophy, psychology, neuroscience, and computer science. Understanding human identity and the science of the mind is no longer a theoretical exercise for Enlightenment coffeeshops or 3 a.m. dorm room existential chats. It is arguably the most urgent research agenda of the next decade, because the internet, the economy, and the law will all have to determine what counts as a person.
“Select all squares with traffic lights.” Sound familiar?
In 2000, Luis von Ahn and a team of Carnegie Mellon University researchers built CAPTCHA: Completely Automated Public Turing test to tell Computers and Humans Apart. CAPTCHA asks a user questions, prompting them to type in letters, identify images in a grid, etc. to ensure that the person is, indeed, a human.
Our humanity is questioned every time we fill out forms, make accounts, reset passwords, etc. There is a longer backstory to CAPTCHA, and why our humanness is being questioned online in the first place.
In October 1950, Alan Turing, a British mathematician, opened an article in Mind with: I propose to consider the question, “Can machines think?” Turing then rejected the question, arguing that the words “machine” and “think” were too ambiguous to survive contact with empiricism. He proposed a substitute — the Turing Test — and the substitute became the standard. The Turing Test suggested that any machine that can sufficiently fool a judge evaluating machine versus human outputs, that machine is intelligent.
76 years later, the LLM revolution obliterated the Turing Test. The question Turing rejected is back, unsolved, and now commercially urgent.
At Roundtable, we suggest the answer — like for many questions in AI — resides in cognitive science. David Marr, a British vision scientist in the mid-20th century, made one of the most enduring arguments in the history of cognitive science: any information-processing system must be understood at three distinct levels. The computational level asks what problem the system is solving and why, the algorithmic level asks what representations and procedures it uses to solve that problem, and the implementational level asks how those procedures are physically realized.
This decomposition becomes powerful when applied to the Turing Test.
The Turing Test operates at the computational level, and only one thin slice of it: input-output mapping. Said simply, the Turing Test only assesses whether the computer can produce intelligent, human-like responses. This has been the source of most serious philosophical critique — Searle’s Chinese Room is the canonical example.
What would be better? A Marr-style three-level Turing Test, which would analyze whether a machine can match human computation and produce human-like responses, but also human algorithm and human implementation as well, breaking down how it solves the problem and physically implements it.
How might a Marr-style three-level Turing test work? By measuring cognitive processing and behavioral biometrics.
Take cognitive processing. This is the domain where psychology and neuroscience have spent decades modeling how humans actually solve problems — under what constraints, with what biases, through what sequences of mental operations. A dominant framework in the field is to treat human cognition as the optimal use of limited resources. Humans don’t have abundant attention: we allocate it selectively and when needed. We overcorrect. Our behavior is shaped by our limitations and optimizations, something my own dissertation argued for cognitive fatigue. These tendencies (measurable and obvious) are the cognitive fingerprints of human thinking.
Next, take behavioral biometrics. This is the implementational fingerprint of that same process: how it manifests in our behaviors, choices, and actions. These include keystroke dynamics, mouse curvature, touch pressure, and scroll velocity – how we are physically interacting with machines.
Together, cognitive processing and behavioral biometrics are the measurable substrate underneath Turing’s test. They are what a process Turing Test would measure.
We at Roundtable call this Proof of Human. The adversarial objection to the Proof of Human research agenda is that this is a losing arms race. We live in an era dominated by the scaling hypothesis: AI capabilities will get smarter and better with more data and more compute. All remaining barriers standing in AI’s way can be barreled down by throwing enough money. Sure, machines cannot fully imitate human processes today, but given enough training data eventually they will.
Even if one rejects the premise that AI capability isn’t humanness, the objection assumes the wrong victory condition. Defense doesn’t need to be permanent; it needs to be economical. The job of an identity primitive – which, said simply, proves who you are – is to make the cost of forgery higher than the value of the forgery. Passwords don’t make accounts unhackable, document verification doesn’t make impersonation impossible, and CAPTCHAs don’t end all spam. Each buys a window reducing instances of hacking – until attackers find a way around them
Roundtable formulates Proof of Human with the same adversarial logic. The question is not whether a capable enough model can imitate a human keystroke pattern — eventually some can. The question is what it costs to do so at scale, against a defender who is also moving. Behavioral biometrics are cheap to measure and expensive to forge convincingly across millions of sessions. Cognitive signatures — how attention degrades, how errors cluster under fatigue, how decisions slow under load — are cheap to elicit and require an attacker to model not just a human, but a particular human, consistently, over time. The arms race is real. The asymmetry is also real.
This is the fear of any defense technology, especially one in the AI era. Analogies can be made with quantum computing and nuclear power, both technoscientific advances that induce arms dynamics. Is building a human verification system a fraught endeavor, or is it something to deeply invest in?
We launched the Proof of Human API in the summer of 2025. The phrase “proof-of-personhood” was becoming increasingly popular in crypto – most popularized by WorldCoin, a Sam Altman-funded orb that scans irises to verify human identity. At Roundtable, we took the underlying concept and rebuilt it on human sciences: psychology, neuroscience, philosophy, and – dare I say – computer science.
For most of history, Turing’s question would be relegated to the academic coffers. Today, it sits underneath every login, signup, transaction, and message we send. Proof of Human won’t definitely solve the philosophical questions — that’s arguably a hard problem. Our ambition is more operational: a minimally viable way to measure reality across any online setting.