I don't have time in the foreseeable future to investigate this, but I would really like someone to look into reliable, cheap, and scalable lie detectors as a cause area. The plausible social benefits in terms of morality, social trust, international trust, and trade could be enormous. But so too could the plausible downsides of enabling dictatorship lock-in and tyranny of the majority.

If you think this line of research is rendered irrelevant by the prospect of transformative AI,^[1] I should also mention the utility of lie detection for credible commitments, verification, and cooperation between states or companies building AI, arms-race de-escalation, and accountability of politicians.

I do advise caution against deferring to me here, but it seems reasonable to believe that (barring existential and political upsets) we'll hit 99%-accurate passenger-car priced lie detector technology within half a century.

Feasibility

A counterargument I've heard is that neuron-level reliable lie detection is infeasible because comparing neural representations of actual beliefs with neural representations of intended meanings, thus detecting a lie, would be incredible hard to pull off because meanings and beliefs are encoded all over the brain.

This argument is based on a false promise that that's how you would design a brain-level lie detector in the first place.

A:[a specific meaning] -> B:[belief/credence] -> C:[verbal representation of meaning and credence]

We're only trying to detect a discrepancy between B and C

Possibilities include

Repeated polygraph testing (bad)
1. If a test is unreliable, it might still be extremely reliable if you just repeat it many times over. But this is premised on how independent the tests are, and this is where all the uncertainty is.
Brain-reading (EEG/fMRI) (maybe)
Invasive BCI or implants (promising!)
Functional Transcranial Doppler (limited)

Repeated polygraph testing (bad)

this might actually be very usefwl for demonstrating truthfwlness for e.g. politicians

I don't mean the traditional polygraph-type lie detectors that look at proxies such as microexpressions, sweating, blood-pressure, and pulse. There's a low upper bound to their accuracy due to the probability of confounders, and most of the social benefit comes from the higher end of reliability.

Brain-reading/EEG/fMRI (maybe?)

Lie detectors using high-level brain-reading,

"Laboratory studies have found rates of accuracy of up to 85%"

Invasive BCI or implants (promising!)

Lie detectors using high-resolution grandmother-based lie detectors will cost more but be more reliable than brain-reading. The tech tree seems different, so progress made on brain-reading may not translate very well (cost of compromise) into progress made on this.

The goal isn't to detect truth, it's to detect whether the subject believes that they are not telling the truth.

A premise for the feasibility of grandmother-based lie detectors is that the relevant representations are sparsely coded. But I think there are strong reasons to expect [links] the brain's neural net to be strongly optimised for sparsity (reduces technical debt, increases cohesion, modularity, robustness, and non-sparse coding is just unfeasible due to weight decay and maybe unknown normalisation stuff).

We have separate areas in the brain for producing and understanding speech, so it doesn't seem infeasible to prevent the detector from falsely going off when you merely contemplate lying or when others speak about it.

You might respond, "we're struggling hard to even interpret artificial neural nets, understanding ANNs is in several ways easier than understanding the brain!"

To which I'd reply that with brain interpretability, we can know what the human is thinking in the first place. At least during the development phase, since test subjects can be asked to lie. This means you have a lot more empirical feedback to constrain theories with.

If we can through neural decoding identify 'grandmother neurons' that reliably activate whenever a concept such as "grandmother" is mentioned, then having it activate when lying doesn't seem too much of a stretch. If lying is encoded differently for different people, then the detector could also be fine-tuned per individual.

Functional Transcranial Doppler (limited)

This is a promising technique (and applicable to AI!), but it's limited to detecting lies for questions that you know the subject has an in-memory answer of.

Recent developments that permit non-invasive monitoring using functional transcranial Doppler (fTCD) technique showed that successful problem-solving employs a discrete knowledge strategy (DKS) that selects neural pathways represented in one hemisphere, while unsuccessful outcome implicates a non-discrete knowledge strategy (nDKS).

Applications to an important subset of deception

This suggests that the DKS model may have a correlate in mnemonic operations. In other words, the DKS model may have a discrete knowledge base (DKB) of essential components needed for task resolution, while for nDKS, DKB is absent and, hence, a "global" or bi-hemispheric search occurs. Based on the latter premise, a 'lie detector' system was designed as described in United States Patent No. 6,390,979.

"this is yet to be tested in forensic practice."

Desirability

If you can reliably convince others that you're honest about your intentions, you are an amazing trading partner. It will be much cheaper to ensure that you're cooperating if they can just verify your intentions at random intervals. Maybe this makes even expensive prototype lie-detector models marketable.

Questions

How much (competent) targeted effort is put into lie detector research already?

Are people already working on this?

It's bizarre that almost all lie detector work has been aimed at eliciting true testimonies from suspected criminals. And furthermore that most work has been done on polygraphs, which are obviously underpowered. [what are implications of this]

Is this another instance of civilizational inadequacy, where research is lacking because people don't have imagination or because it's bad reputation or because people aren't backward-chaining?

Most people aspire to hide their selfish motivations by consciously or subconsciously feigning altruistic behaviour. And humans self-deceive in order to better be able to convince others that they truly have those altruistic motivations (Trivers-like). If motivations for actions are unconcealable, people either turn to overt selfish behaviour, or they self-deceive to credibly hold altruistic motivations for so long that it just becomes their motivation.

If there are no opportunities to actually do the selfish behaviour (e.g. trying to convince someone of X in order to seem like you know stuff, by citing studies you know to be flawed), and you have to self-deceive in order to preserve ability to do the selfish thing when opportunity comes, and self-deception is more computationally expensive than straightforward motivation, you might just eventually learn intrinsic motivation to do the behaviour in the honestly altruistic way (e.g. only trying to convince someone of X when you have sufficient evidence, and not trying to appear more confident than warranted).

What happens if lie-detectors are cheap and universal? And what happens if people willingly use them (which they might feel pressured to do if their friends do)?

Either they converge on consistently selfish behaviour (since there's no profit in faking altruism when it's unlikely you'll pull it off), or consistently altruistic behaviour (no profit in trying to be secretly selfish when it's hard to pull off). Like Schrödinger motivations? They settle on consistently and altruistic or selfish once reliably observed?

This is kinda what happens in the rationalist/EA community with some people: there's a common culture for more honesty, which means people honestly report their self-centered motivations and aren't judged for it, and also have more honest altruistic behaviour (I think) compared to less-honest cultures.

Update 7/11/2022, resources:

https://forum.effectivealtruism.org/posts/qfDeCGxBTFhJANAWm/a-new-x-risk-factor-brain-computer-interfaces-1
- "Lie detection via neurotechnology, or mind-reading complex thoughts, seems quite difficult, and faces structural issues in that the representations for complex thoughts are going to be developed idiosyncratically in each individual, whereas things like optic nerve connections and the lower levels of V1 can be tracked by their definite inputs and outputs, shared across humans"
  -- Carl Shulman
- ^ The above seems wrong for the reason that you don't need to identify how individual memories/concepts are stored, you just need to identify a common activation pattern for "lying" is activated (seems likely to exist), and/or a mismatch between B (belief-strength/credence) and C (verbal confidence).
What people don't get is that the main benefits from lie detection comes from people willingly using them in order to credibly signal their honest intentions.
- Bad Nash equilibria, Braess paradox, Nash attack, Price of Anarchy

^{^}
These are ad hoc reasons for lie detectors. I was excited about lie detectors for reasons independent of AI. If you're primarily concerned with AI, I highly doubt you should be studying human lie detectors.