In principle, machine learning approaches could produce lie detectors far more reliable than the polygraph. If this happens soon, it will matter a great deal: reliable lie detection could make loyalty purges much more effective, create a useful check on political and AI-lab leaders, make international treaties more robust, and make authoritarian regimes more stable.
I argue the balance between these effects depends largely on cooperation cost: how much willing participation a detector needs from its subject. In particular, lie detection technology with extremely low cooperation cost is better to accelerate than lie detection technology with higher cooperation cost. I remain agnostic here about whether we should, in general, accelerate lie detector technology—this very well might be net-negative. But if we do, we should accelerate low-cooperation-cost lie detection and ensure it is widely available.
In this article, I’ll describe cooperation cost and argue that remote lie detectors strike a better balance of benefits and drawbacks. Then, I’ll explore the most important effects in more depth and see how they differ for remote and in-person lie detectors. Lastly, I’ll consolidate this exploration and touch on some other factors.
Cooperation cost
Lie detector technology could operate on many different input modalities, some of which require more willing participation from the subject. For example, lie detection on written text has very little cooperation cost; it only requires the subject to send writing or have publicly available writing. Conversely, lie detection based on an invasive neural device has extremely high cooperation cost. From lowest to highest cooperation cost, lie detectors could operate on:
Written text
Audiovisual data
Polygraph-like biometric data
Neural data from a wearable, mobile device
Neural data from a fixed device (e.g., fMRI)
Neural data from an invasive device, requiring surgery
(and many other things and combinations of things)
There is a significant contrast in cooperation cost between lie detectors that require the subject to be present in person, specifically for the purposes of a lie detection test (i.e., polygraph-like biometric data and beyond), and lie detectors that can operate remotely and asynchronously based on data not particularly related to lie detection (i.e., written text and audiovisual data). In the rest of this article, I’ll focus on the difference between remote and in-person lie detectors.
Remote is better
One framing of lie detectors is that they tend to do harm when the powerful point them at the weak, and good when the weak point them at the powerful; remote detectors help more with the latter. More powerful people can more easily coerce others and are also harder to be coerced themselves. So when lie detectors require the subject to agree to an interview and go somewhere in person, the powerful can use them on the less powerful, but the reverse is very hard. Conversely, when lie detectors are remote and require no cooperation, consent, or even awareness, anyone can use them on anyone else. The balance of good effects to bad effects is better with remote lie detectors.
Price and availability also matter, but remote detectors are more likely to be cheap and easy to release, while in-person detectors are more likely to be expensive and easy to withhold. In-person detectors that rely on specialized biometric or neural-data hardware will be much more expensive than remote algorithms that can run on a small amount of compute. A remote algorithm can be readily open-sourced and run by anyone who knows to ask an AI assistant how to use it, but even if in-person detectors have completely open design and specifications, purchasing or attempting to replicate the hardware will be much more difficult. Remote lie detectors could still be developed privately and kept proprietary. It is essential for the benefits I discuss below that they are made widely available. But their form factor (software) makes them easy to widely release, whereas in-person detectors’ form factor makes them easy to withhold from the public.
In the next few sections of this article, I’ll explore the effects of lie detectors more concretely, especially as they relate to cooperation cost.
Loyalty screening and purges
The US government already uses polygraphs to screen and investigate its own personnel, and lie detectors that actually work would likely incentivize more testing and make existing testing significantly more effective. In the worst case, lie detectors could be used for strong loyalty checks across government, exacerbating the risk of concentration of power.
Loyalty screening could also extend to frontier AI labs, even in the near term. The Employee Polygraph Protection Act covers essentially any kind of lie detector, so labs probably could not conduct any kind of screening in a private capacity in the near term. However, the act contains a carveout for the government to screen employees of DoD contractors, so the USG could screen employees of OpenAI, Google, xAI, and Anthropic. This would probably be used in a national security context to oust foreign spies, but in the worst case, could be used to oust whistleblowers or ten safety-concerned people on the inside.
Importantly, all of these checks occur in settings where high cooperation costs aren't a barrier, so they are equally boosted by remote and in-person lie detector acceleration.
Public checks on leaders
Accelerating lie detection could allow the public to select better political and frontier lab leaders, reducing concentration-of-power risk. If frontier lab CEOs submit to lie detector tests and answer questions about, for example, what they plan to do with superintelligence or if they plan to install secret loyalties, this could help rule out certain extremely negative tail risks. If politicians submit to lie detector tests, this could reduce the chance they mislead the public or fail to follow through on campaign promises, which could be broadly positive. Political lie detector tests could also help rule out some tail risks related to power consolidation.
Unlike loyalty screening, threat models differ significantly here between remote and in-person lie detectors.
Theoretically, in-person lie detectors could improve leader selection, but I expect this to be very unlikely, especially in the near term. Candidates in elections that publicly submit to lie detector tests might outcompete those who don’t, so an equilibrium could emerge in which all candidates willingly submit to tests. With significant public pressure, lab leaders may also submit to lie detector interviews. In practice, however, I expect leaders not to want to submit to lie detector exams, and to come up with many credible-seeming excuses, spread disinformation about lie detector accuracy, politicize lie detectors, and otherwise get out of them easily, with relatively little downside. I think this is especially likely in the near term, when there is no common knowledge about lie detectors' robustness and accuracy (perhaps because they are not yet extremely robust and accurate).
With remote lie detectors, these problems don't prevent us from performing lie detection on politicians and lab leaders. For example, we could run audiovisual lie detectors on videos of public interviews, speeches, and debates. Audiovisual lie detectors require so little cooperation that they are even robust to certain adversarial responses (decreasing public appearances, feeding nonanswers to certain questions, etc.), because they can be run on historical videos. For example, we have videos of lab leaders discussing concentration of power and the undesirability of unilateral control of AI, so we might be able to rule out certain tail risks even if leaders start substantially controlling all of their public appearances.
Moreover, the open use of remote lie detectors on leaders (assuming they do work) could provide a public proving ground and an on-ramp for more intensive use, further reducing the scope for evasion. If lie detectors become highly visible and show promising initial results, the public could start demanding their use in debates or more targeted interviews.
I expect better political leader selection to be positive on average, but not necessarily; for example, if the more honest candidate in an election supports worse policies, but is elected for being more honest, lie detectors might be net negative. I expect the results of lab leader screening to be positive with more confidence.
International treaties
Lie detection could help verify international AI treaties, for example, as an added check against covert AI projects. Beyond AI treaties, lie detectors could broadly improve international coordination. For example, they could similarly create a verification mechanism for bioweapons treaties or help ensure that prisoner lists and weapons declarations in peace agreements are complete. They could also vet people who collaborate on high-stakes international projects; for example, a country might accept more intrusive arms-control inspections if it could confirm the inspectors are not spying for another government.
Near-term acceleration could help create common knowledge of a working technology, making lie detectors a more salient and likely choice for future international coordination. That said, lie detectors created in the near term probably will not be immediately suited for international treaty verification because, for example, the Chinese government would not trust a technology developed by a US startup rather than one jointly developed for a treaty.
Like loyalty screens, these checks occur in settings where high cooperation costs are not a barrier, so they are equally boosted by remote and in-person lie detector acceleration.
Authoritarian regimes
One worry is that accelerating lie detectors might make authoritarian regimes more repressive. Remote lie detectors could even enable mass monitoring, without consent or even awareness, to a greater extent and more cheaply than in-person lie detectors. However, I think this worry is less important than it sounds.
It matters less whether lie detectors create a more repressive regime than whether they stabilize and enable the propagation of bad values into the far future. It is fairly unclear, and implementation-specific, how much the near-term acceleration of lie detectors would have a counterfactual stabilizing effect. Near-term acceleration could make the technology more salient and bring it about in authoritarian regimes sooner, but if lie detection in any form proves useful for authoritarian regimes, it is probably overdetermined that they will develop it. As such, I think the broader effects of accelerating lie detectors on authoritarian regimes are negative but somewhat unclear.
Authoritarian regimes are susceptible to significant regime or value change via two main channels: public revolts and coups.
Lie detection might bolster public monitoring and purging but will likely have little counterfactual effect on stability, especially via remote versions of the technology. The CCP (along with many other modern regimes) is already extremely robust to revolts because of elaborate party co-optation schemes, controlled repression efforts, and extreme modern military power. Moreover, there aren’t many clear places where remote lie detection could fit in a way that strongly benefits the regime. For example, running lie detectors on surveillance cameras would be relatively pointless; activities central to a meaningful revolt effort would already have to avoid prosaic monitoring, so adding a layer of lie detection would have little counterfactual impact on stability. Screening, for example, security at the time of appointment similarly might add little over existing tools, because sincere answers to hypothetical questions (e.g., “Yes, I would shoot”) poorly predict behavior under extreme pressure. Lie detection could be worrying alongside other AI-enabled surveillance, such as analyzing and consolidating disparate records and data, flagging individuals as suspicious, and then inviting them to an in-person lie-detector interview. However, regimes already have fairly workable machinery for this, so lie detectors probably wouldn’t have massive counterfactual effects. Even if AI progress is destabilizing enough to make a regime more susceptible to revolts, lie detection will be counterfactually responsible for a regime staying in power against a public revolt only in a very narrow class of scenarios, and the most important uses are in settings where high cooperation cost is not a barrier.
Lie detection probably has a larger effect on the insider channel (coups, elite constraint). Coups rarely produce democracy, but they are the main way that values change inside a consolidated autocracy. Routinely asking regime insiders about current intentions ("Are you planning anything now? Do you think the leader is legitimate?") might stabilize the regime. It could also weaken the loyalty-competence trade-off: dictators sometimes hire less competent subordinates because poor outside options keep them loyal, but a dictator who can monitor insiders’ intentions directly and punish disloyalty could afford to hire competent ones. That would make the regime both more stable and more capable. On the other hand, transparent lie detection could solve the collective action problem for the dictator's allies, making a coup more likely.
Takeaways
Some of the most important effects of near-term acceleration of lie detectors might be:
increasing the extent and efficacy of loyalty screening and purges.
creating more robust public checks on lab and political leaders.
increasing the chance lie detectors are used to enable international coordination.
stabilizing authoritarian regimes.
I expect (1) to exacerbate concentration of power and be largely net-negative, but I expect (2) to attenuate the concentration of power, or at least decrease the chance power is concentrated in the wrong hands, and be largely net-positive. (2) is significantly easier with remote lie detectors, but (1), (3), and (4) are equally enabled by both remote and in-person lie detectors.
Thank you to Linchuan Zhang, Dylan Bowman, Anastasia Wei, Zephaniah Roe, and others for helpful writing, discussion, and feedback. All opinions are my own.
In principle, machine learning approaches could produce lie detectors far more reliable than the polygraph. If this happens soon, it will matter a great deal: reliable lie detection could make loyalty purges much more effective, create a useful check on political and AI-lab leaders, make international treaties more robust, and make authoritarian regimes more stable.
I argue the balance between these effects depends largely on cooperation cost: how much willing participation a detector needs from its subject. In particular, lie detection technology with extremely low cooperation cost is better to accelerate than lie detection technology with higher cooperation cost. I remain agnostic here about whether we should, in general, accelerate lie detector technology—this very well might be net-negative. But if we do, we should accelerate low-cooperation-cost lie detection and ensure it is widely available.
In this article, I’ll describe cooperation cost and argue that remote lie detectors strike a better balance of benefits and drawbacks. Then, I’ll explore the most important effects in more depth and see how they differ for remote and in-person lie detectors. Lastly, I’ll consolidate this exploration and touch on some other factors.
Cooperation cost
Lie detector technology could operate on many different input modalities, some of which require more willing participation from the subject. For example, lie detection on written text has very little cooperation cost; it only requires the subject to send writing or have publicly available writing. Conversely, lie detection based on an invasive neural device has extremely high cooperation cost. From lowest to highest cooperation cost, lie detectors could operate on:
There is a significant contrast in cooperation cost between lie detectors that require the subject to be present in person, specifically for the purposes of a lie detection test (i.e., polygraph-like biometric data and beyond), and lie detectors that can operate remotely and asynchronously based on data not particularly related to lie detection (i.e., written text and audiovisual data). In the rest of this article, I’ll focus on the difference between remote and in-person lie detectors.
Remote is better
One framing of lie detectors is that they tend to do harm when the powerful point them at the weak, and good when the weak point them at the powerful; remote detectors help more with the latter. More powerful people can more easily coerce others and are also harder to be coerced themselves. So when lie detectors require the subject to agree to an interview and go somewhere in person, the powerful can use them on the less powerful, but the reverse is very hard. Conversely, when lie detectors are remote and require no cooperation, consent, or even awareness, anyone can use them on anyone else. The balance of good effects to bad effects is better with remote lie detectors.
Price and availability also matter, but remote detectors are more likely to be cheap and easy to release, while in-person detectors are more likely to be expensive and easy to withhold. In-person detectors that rely on specialized biometric or neural-data hardware will be much more expensive than remote algorithms that can run on a small amount of compute. A remote algorithm can be readily open-sourced and run by anyone who knows to ask an AI assistant how to use it, but even if in-person detectors have completely open design and specifications, purchasing or attempting to replicate the hardware will be much more difficult. Remote lie detectors could still be developed privately and kept proprietary. It is essential for the benefits I discuss below that they are made widely available. But their form factor (software) makes them easy to widely release, whereas in-person detectors’ form factor makes them easy to withhold from the public.
In the next few sections of this article, I’ll explore the effects of lie detectors more concretely, especially as they relate to cooperation cost.
Loyalty screening and purges
The US government already uses polygraphs to screen and investigate its own personnel, and lie detectors that actually work would likely incentivize more testing and make existing testing significantly more effective. In the worst case, lie detectors could be used for strong loyalty checks across government, exacerbating the risk of concentration of power.
Loyalty screening could also extend to frontier AI labs, even in the near term. The Employee Polygraph Protection Act covers essentially any kind of lie detector, so labs probably could not conduct any kind of screening in a private capacity in the near term. However, the act contains a carveout for the government to screen employees of DoD contractors, so the USG could screen employees of OpenAI, Google, xAI, and Anthropic. This would probably be used in a national security context to oust foreign spies, but in the worst case, could be used to oust whistleblowers or ten safety-concerned people on the inside.
Importantly, all of these checks occur in settings where high cooperation costs aren't a barrier, so they are equally boosted by remote and in-person lie detector acceleration.
Public checks on leaders
Accelerating lie detection could allow the public to select better political and frontier lab leaders, reducing concentration-of-power risk. If frontier lab CEOs submit to lie detector tests and answer questions about, for example, what they plan to do with superintelligence or if they plan to install secret loyalties, this could help rule out certain extremely negative tail risks. If politicians submit to lie detector tests, this could reduce the chance they mislead the public or fail to follow through on campaign promises, which could be broadly positive. Political lie detector tests could also help rule out some tail risks related to power consolidation.
Unlike loyalty screening, threat models differ significantly here between remote and in-person lie detectors.
Theoretically, in-person lie detectors could improve leader selection, but I expect this to be very unlikely, especially in the near term. Candidates in elections that publicly submit to lie detector tests might outcompete those who don’t, so an equilibrium could emerge in which all candidates willingly submit to tests. With significant public pressure, lab leaders may also submit to lie detector interviews. In practice, however, I expect leaders not to want to submit to lie detector exams, and to come up with many credible-seeming excuses, spread disinformation about lie detector accuracy, politicize lie detectors, and otherwise get out of them easily, with relatively little downside. I think this is especially likely in the near term, when there is no common knowledge about lie detectors' robustness and accuracy (perhaps because they are not yet extremely robust and accurate).
With remote lie detectors, these problems don't prevent us from performing lie detection on politicians and lab leaders. For example, we could run audiovisual lie detectors on videos of public interviews, speeches, and debates. Audiovisual lie detectors require so little cooperation that they are even robust to certain adversarial responses (decreasing public appearances, feeding nonanswers to certain questions, etc.), because they can be run on historical videos. For example, we have videos of lab leaders discussing concentration of power and the undesirability of unilateral control of AI, so we might be able to rule out certain tail risks even if leaders start substantially controlling all of their public appearances.
Moreover, the open use of remote lie detectors on leaders (assuming they do work) could provide a public proving ground and an on-ramp for more intensive use, further reducing the scope for evasion. If lie detectors become highly visible and show promising initial results, the public could start demanding their use in debates or more targeted interviews.
I expect better political leader selection to be positive on average, but not necessarily; for example, if the more honest candidate in an election supports worse policies, but is elected for being more honest, lie detectors might be net negative. I expect the results of lab leader screening to be positive with more confidence.
International treaties
Lie detection could help verify international AI treaties, for example, as an added check against covert AI projects. Beyond AI treaties, lie detectors could broadly improve international coordination. For example, they could similarly create a verification mechanism for bioweapons treaties or help ensure that prisoner lists and weapons declarations in peace agreements are complete. They could also vet people who collaborate on high-stakes international projects; for example, a country might accept more intrusive arms-control inspections if it could confirm the inspectors are not spying for another government.
Near-term acceleration could help create common knowledge of a working technology, making lie detectors a more salient and likely choice for future international coordination. That said, lie detectors created in the near term probably will not be immediately suited for international treaty verification because, for example, the Chinese government would not trust a technology developed by a US startup rather than one jointly developed for a treaty.
Like loyalty screens, these checks occur in settings where high cooperation costs are not a barrier, so they are equally boosted by remote and in-person lie detector acceleration.
Authoritarian regimes
One worry is that accelerating lie detectors might make authoritarian regimes more repressive. Remote lie detectors could even enable mass monitoring, without consent or even awareness, to a greater extent and more cheaply than in-person lie detectors. However, I think this worry is less important than it sounds.
It matters less whether lie detectors create a more repressive regime than whether they stabilize and enable the propagation of bad values into the far future. It is fairly unclear, and implementation-specific, how much the near-term acceleration of lie detectors would have a counterfactual stabilizing effect. Near-term acceleration could make the technology more salient and bring it about in authoritarian regimes sooner, but if lie detection in any form proves useful for authoritarian regimes, it is probably overdetermined that they will develop it. As such, I think the broader effects of accelerating lie detectors on authoritarian regimes are negative but somewhat unclear.
Authoritarian regimes are susceptible to significant regime or value change via two main channels: public revolts and coups.
Lie detection might bolster public monitoring and purging but will likely have little counterfactual effect on stability, especially via remote versions of the technology. The CCP (along with many other modern regimes) is already extremely robust to revolts because of elaborate party co-optation schemes, controlled repression efforts, and extreme modern military power. Moreover, there aren’t many clear places where remote lie detection could fit in a way that strongly benefits the regime. For example, running lie detectors on surveillance cameras would be relatively pointless; activities central to a meaningful revolt effort would already have to avoid prosaic monitoring, so adding a layer of lie detection would have little counterfactual impact on stability. Screening, for example, security at the time of appointment similarly might add little over existing tools, because sincere answers to hypothetical questions (e.g., “Yes, I would shoot”) poorly predict behavior under extreme pressure. Lie detection could be worrying alongside other AI-enabled surveillance, such as analyzing and consolidating disparate records and data, flagging individuals as suspicious, and then inviting them to an in-person lie-detector interview. However, regimes already have fairly workable machinery for this, so lie detectors probably wouldn’t have massive counterfactual effects. Even if AI progress is destabilizing enough to make a regime more susceptible to revolts, lie detection will be counterfactually responsible for a regime staying in power against a public revolt only in a very narrow class of scenarios, and the most important uses are in settings where high cooperation cost is not a barrier.
Lie detection probably has a larger effect on the insider channel (coups, elite constraint). Coups rarely produce democracy, but they are the main way that values change inside a consolidated autocracy. Routinely asking regime insiders about current intentions ("Are you planning anything now? Do you think the leader is legitimate?") might stabilize the regime. It could also weaken the loyalty-competence trade-off: dictators sometimes hire less competent subordinates because poor outside options keep them loyal, but a dictator who can monitor insiders’ intentions directly and punish disloyalty could afford to hire competent ones. That would make the regime both more stable and more capable. On the other hand, transparent lie detection could solve the collective action problem for the dictator's allies, making a coup more likely.
Takeaways
Some of the most important effects of near-term acceleration of lie detectors might be:
I expect (1) to exacerbate concentration of power and be largely net-negative, but I expect (2) to attenuate the concentration of power, or at least decrease the chance power is concentrated in the wrong hands, and be largely net-positive. (2) is significantly easier with remote lie detectors, but (1), (3), and (4) are equally enabled by both remote and in-person lie detectors.
Thank you to Linchuan Zhang, Dylan Bowman, Anastasia Wei, Zephaniah Roe, and others for helpful writing, discussion, and feedback. All opinions are my own.