The world is now a state where AI is sophisticated enough that it can read and meaningfully interpret draft bills/legislation.

Human politicians are replaced with AIs in an attempt to combat corruption, closed-door deals etc that are currently commonplace.

The citizens of this society give more/less power to each AI based on votes.

Other functions of a government are still left to humans, however passing of new legislation or changes to existing legislation can only be done by the AIs.

The AIs

The AIs would be open-source algorithms that are run in pre-determined environments (e.g. a docker image).


  • The draft bill/legislation.
  • It also has full read access to any publicly available information (i.e. the internet), and also any private information currently available to governments.


  • It would keep extensive logs (write-only) that could be used for analysis and criticism (and ultimately as well as development of new AI's).
  • The final output would be a simple Yes/No - i.e. it does or does not approve the legislation.

For safety it cannot interact with the internet, i.e. its internet access is read-only.

How to approve legislation

  • Anyone can produce a bill/draft legislation.
  • This is then fed on to voting by the AI's.
  • More than 50% of the voting power (i.e. each AI's decision x number of supporters) must be in favour of the legislation for it to pass.

Voting Rules

  • The human citizens of the state will vote for AIs once per political term.
  • The AI's are used for the entire political term.
  • Every human over voting age and fit to vote is required to vote.
  • Each vote has equal weight but can be divided among multiple AI's.

This is an extension of another discussion on preventing corruption, and specifically one answer that used robots.

I would like your help in finding how the system:

  • Could prevent corruption
  • Could efficiently develop the state's laws? (timely response to events)
  • Could be abused by humans
  • Could malfunction (does not serve any human a benefit while being to the detriment of society)
  • Could evolve into a more sinister political system.
  • Would be less accurate than a human system. (Where humans would better represent the interests of their voters)
  • Might need a killswitch? (Some western political systems give executive powers to one person who can overturn a decision/government)
  • Might be technically possible.

And finally, the campaigns - how would people decide which AI to vote for? One suggestion is that you could run a series of bills/legislation drafts through the algorithm and determine it's tendencies (like a system test)

  • 1
    $\begingroup$ If the algorithms are openly available, what's to prevent a person from putting just enough items in a bill to get it passed while tacking on whatever other thing they want to actually pass through. Not that this isn't common already in some governments (cough US congress), but this seems like a more sure fire way. $\endgroup$ Commented Mar 28, 2016 at 6:23
  • $\begingroup$ Thanks Chis - and I'm sure most governments would do similar. So do you think we would be any worse off under an AI system? $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 6:53
  • $\begingroup$ with human governments at least people are a bit harder to crack and can't be simulated. If this AI system can pass laws quickly (even when not an emergency) then I can see it being abuswd and worse. $\endgroup$ Commented Mar 28, 2016 at 16:12
  • $\begingroup$ Its a cool idea, you could actually test how each AI would vote under different conditions as part of the election and the AI can't go back on its promises $\endgroup$ Commented Mar 28, 2016 at 23:18
  • 2
    $\begingroup$ First sanity check: what stops an overlooked mistake in the code from causing the AIs to add new legislature changing the voting rules so the humans no longer get a vote? Second sanity check: if you rely on a killswitch approach to resolve the first sanity check, how do you avoid corruption in the hands of those who can use the killswitch? If you're not careful, you may concentrate power/corruption rather than diffusing it. $\endgroup$
    – Cort Ammon
    Commented Mar 29, 2016 at 2:04

6 Answers 6


Although possible from a technical standpoint, your idea runs into a few problems.

An AI is tremendously complicated to create. Complicated enough that you would have to reuse much of the old candidate's code in order to have the next ready for the election. At that point, you're just updating the version of your candidate for the most recent issues. Can that really be considered a "different" AI? Perhaps that doesn't matter for your world, but it could bring up some interesting plot elements ("RedBot wins Senate race again!"). A related issue is that an AI of this capacity would also be expensive to develop and maintain. Expensive enough that only certain large organizations (let's call them "parties") would have the available funds to reliable field a new AI for each election. Other small entities (let's call them "independents") would occasionally have enough local support to get elected, but it would be a rarity.

Another issue is maintenance. There are two alternatives to this. First, you can take the AI as-is at the time of election and not allow any updates during the term. 6 years (for a Senate term) is a long time for bugs to go unresolved. If your Senate is mistakenly routing billions of dollars into the Save-The-Nyancats fund, there is nothing you can do about it. The alternative is allowing updates during the term. However, the problem here is that you could essentially fix votes as if the party was actually in control. For example, the Senator-AI doesn't vote the way the party wanted, so it gets "fixed" as part of a "bug". This would be very difficult to keep track of, as the AI software is immense.

Even after all that, your Senator-AI can still be manipulated . Your AI falls into one of two categories, weak or strong. A weak AI is more bound by its creator's parameters, while a strong AI can learn independently and for our purposes is sentient. So if your AI is weak, it could be essentially hard-wired to do whatever the party tells it to (ex: RedBots are programmed to reject all proposals from BlueBots). On the other hand, if the AI's are strong, then...maybe? However, if any part of your laws violate AI rights (killswitch, external bug fixing, etc.), they would immediately be thrown out the window as the AIs are sentient enough to know their rights. All hail the Singularity Party.

As a last point, I think that open source politicians would be a terrible mistake. Imagine if a political party could have a private copy of the Senate that they feed bills until they find one that passes. This could be abused very easily.

The average voter has little knowledge of programming, let alone complex AI. Those who could take the most advantage of the open source code would be the opposition looking for weaknesses. Now, it might be more beneficial to have a way for the public to interact with the AI (a "Talk to your Senator" app if you will), which would go a lot further in terms of public trust. Just look at today's campaigns. Most voters are interested in the general policies and overall character of the candidates rather than what minutiae of legislature they are willing to pass.

Some alternatives

Strong AIs as regular candidates
If the AIs are essentially sentient, why not have them run against human candidates. Let the people decide who they think is better. While this removes a lot of the efficiency of an all AI government, it also allows for a failsafe in case the AIs are not performing as they should.

One AI with voting algorithms
The "AI based on votes" phrase reminded me of a completely different concept, neural networking. With this setup, voters would vote directly on general political issues rather than for a particular candidate, rather like a semi-annual referendum. There would be one AI that would be "trained" by these votes, so that it's decision-making was directly tied to the people's will. This is not perfect by any means, and would override the existing political system, but it would be less susceptible to the machinations of political parties.

  • $\begingroup$ I think sentient AI could simply become the ruler of the world, and not just through legislation - it could as well bypass the "read-only" characteristic of its connection to the Internet by URL address injections or by hacking it's way through it. $\endgroup$ Commented Mar 28, 2016 at 20:16
  • $\begingroup$ kys you bring up some really good points - especially around funding of development work (which would also fund campaign efforts). I also like your idea on introducing strong AIs as regular candidates as opposed to an all-AI candidature. $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 23:30
  • $\begingroup$ However I disagree that open source politicians would be a terrible mistake - assuming that the AIs do have the best interest of their voters (a very big, perhaps impossible, if) at hand, all bills that pass would intrinsically be OK with their voters - so no harm done, right? $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 23:30
  • $\begingroup$ added a bit to the open source part. I'm assuming your main intent with having the AI be open source is public trust, $\endgroup$
    – Kys
    Commented Mar 29, 2016 at 13:40
  • $\begingroup$ I guess the main intent is to make the system serve the voters interests which should in turn increase public trust? Perhaps de-humanising the legislative duties would make voters care about the legislation tendencies vs voting for the most charismatic politician? $\endgroup$
    – timyhac
    Commented Mar 30, 2016 at 0:40


How do I know who (what AI) to vote for? Test vs code review

I could either feed each AI it test scenarios and see if I agree with its response, which is an awesome ability or I examine the open source code directly.

There are two problems I want to spot, deliberate design decisions in the AI I disagree with, say tax breaks for its parent company, and accidental errors in its though process that will produce bad votes.

For test the space of all possible legislation and all possible states of its information source (the internet) are so massive it is very had to ensure good test coverage. A back door could be built in that passes all legislation whose name appears on a certain obscure website and I would never find it.

Code review is hard

If you give a good programmer a thousand lines of badly written or obscured code he or she will have trouble understanding exactly what is happening and the complexity grows exponentially with the size of the program. A complex AI would be hard to understand even if it was designed to be easy to understand, and if it was designed to have some special interest, it would be hard to spot.

We might not understand our own laws Current laws have to be simple enough for politicians to understand and defend to their constituents. AI drafted legislation could be arbitrarily complex and so if it was bad we might not ever figure it out.

Things move to fast

One of the problems with automated trading is that computers are fast. If the computers decide that a stock is going down they will sell it so fast that its value will 90% in 5 minutes, since this happed limits on high speed trading have been introduced. The equivalent problem is if the AIs have some crazy ideas they can pass hundreds of laws before humans could intervene

  • $\begingroup$ Brilliant points - other commenters have alluded to these deliberate design decisions that an AI could pass arbritary bill via a magic code in the bill that would be impossible to test against. $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 23:33
  • $\begingroup$ Things move too fast - also a brilliant point. Do you think there is any way that we could retain the efficiency of AI politicians but also prevent the things moving too fast condundrum? $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 23:34
  • $\begingroup$ @timyha To fast is the simplest problem. Humans have to enforce the laws so they could be written and passed instantly but it would be a lag of days, weeks or months before humans get them enforced. $\endgroup$ Commented Mar 28, 2016 at 23:40

When you are just replacing the legislative with AIs, it might be possible to get just laws. Assuming that we actually might manage to create 'just AIs', of course (the other answers have a lot of ideas of fail-points like who programs them, how are they maintained, how to actually program a Republican vs. a Democrat AI etc.).

So, assuming we now have absolutely just AIs that can't be corrupted at all. Points where corruption might be introduced:

  • Who is allowed to suggest new laws? If everyone can suggest new laws, the AIs get flooded by truckloads of garbage because of course everyone will try out creating their own law. Like, you have to wear socks of the same color when using a public bathroom. So, depending on the AIs, you either get so many new (nonsense) laws that no human will ever have a chance anymore of just reading them in their lifetime, or you need to introduce an algorithm that weeds out 'unnecessary' laws -> one more point where corruption can be introduced. If only some people can suggest new laws, I think the time-consuming part of creating bills won't be the passing of a bill (5 seconds wait and then a 'yes' or 'no') but the formulating of a bill. Also, corruption can begin with a necessary law just never being suggested to the legislative (no loopholes or other finery necessary).
  • the entire judicative (judge + jury) which is responsible for turning laws into practice, is still human. And interpretation is 90% of the law. Once you've got some precedences out of the way, the spirit of a law might be pretty much circumvented completely. And humans are damnably good at finding and/or creating loopholes with sufficient motivation (i.e. bribe money)

So, having an AI legislative probably won't help a lot with the corruption problem. Neither will it help with the timeliness problem because people will now fight over the law text outside of senate instead of inside.

Once you start replacing the two critical points with other AIs though, your acceptance will go down the drain (do humans really want to let themselves be judged by robots? to let robots suggest their new laws?).

And last but not least -- interpreting laws (i.e. using already defined rules) is something completely different from adding new laws to the catalogue. How does an AI know which new laws are 'good' and which are 'bad', i.e. when to say 'yes' or 'no'? Public voting? Skimming the general opinion of hundreds and thousands of internet discussion threads (meaning being fed with a lot of extremist views)?

  • $\begingroup$ In this scenario, we are assuming that the AIs are just (with respect to their voters) so the last two paragraphs of this answer I won't discuss. But I guess the two dot points you bring up lead me to question what the point of legislation is at all. To me law is a mechanism to encourage certain behaviours, and also to provide compensation for victims of 'bad' behaviour. So there is no point to having laws unless it can be used for these purposes. This is a trend that seems to be emerging from other people's answers too. $\endgroup$
    – timyhac
    Commented Mar 31, 2016 at 0:26
  • $\begingroup$ @timyha I just wanted to point out that using AIs can't really solve the 'human sickness', i.e. people trying to get the most gain for themselves, prejudice, etc. It will just migrate to different venues. Will there be an over-all benefit to having a just AI bill-passing system? Maybe. But it will never be a miracle-cure for corruption etc. $\endgroup$
    – subrunner
    Commented Mar 31, 2016 at 7:35

Could prevent corruption

Yes, the system in place would be mostly corruption-free in the legislation. But this does not mean that the whole functioning of the government would be transparent. Mostly laws contain a gray zone where a significant amount of decision making is left at the judgement of the official in charge. However, that is not as damning as the corruption of people in higher chambers.

Could efficiently develop the state's laws? (timely response to events)

Depends completely on the way it is meant to function. Humans are very flexible and have the ability to react quickly or slowly, depending on the nature of the problem. Whether your AI would function quickly or slowly, relies on which instructions it follows under which conditions.

Yes, you can modify the software in order to enable it to make quick decisions, but just as quick decisions made by humans, these would be susceptible to selection bias (of the programmers) and insufficient data.

Could be abused by humans

Anything made humans can be abused by humans. Witty programmers can put in logical loopholes and traps even in open source software. Remember the heart bleed vulnerability?

It is only a matter of the caliber of the programmers aimed at putting loopholes and vulnerabilities in the software versus the programmers aimed at removing these menaces. The smarter ones win.

Could malfunction (does not serve any human a benefit while being to the detriment of society)

Depends completely on the program coding and the hardware it is running at. Most software developed, does malfunction and misbehave once in a while. This is why we get upgrades and newer versions.

A perfect software (which never malfunctions) is either too simple to host any loopholes or must be made by the perfectly logical and intelligent programmer(s) (who do not exist).

Could evolve into a more sinister political system.

Depends on what you call sinister and also depends on the the internal coding of the intelligent part of the AI. It also depends heavily on the type of society it is legislating for. For example, if it presented with data that indicates a constant rise in violent crimes and corruption, the logical decision (logical is a highly subjective term btw) would be use more force to curb these menaces. So all in all, it depends on its coding and the type of society.

Would be less accurate than a human system. (Where humans would better represent the interests of their voters)

Depends on how intelligent the software is. If you have high caliber political experts, human psychology experts, legal experts and programming experts in the panel, then you would probably get a better system than a human run government.

But even smart people make errors. And programming related errors are the most deadly of all when considering the behavior of an AI.

Let it suffice to say that the software would give better results where fairness and transparency is the major concern and wold give poorer results where wisdom and understanding of the society is the key factor.

Might need a killswitch? (Some western political systems give executive powers to one person who can overturn a decision/government)

Depends on your government system/constitution. The constitution is written and devised by humans, not AI. So it is completely on your own choice to put a kill switch for the system or not.

Might be technically possible.

Yes it is technically possible to build a governance system like that with our current technology. However, since it would be a completely technical government, there are major chances its behavior and responses would be highly predictable, which is not a good thing.

  • $\begingroup$ Thanks for your quick response @youstay-igo. I come from a software product management background and I believe that any software is only as good as its testing (system testing, unit testing, etc).Do you think it is possible to come up with some tests that could circumvent human abuse attempts? $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 6:45
  • $\begingroup$ By sinister I mean something that starts out to solve one problem but ends up causing different and possibly worse problems (i.e. 1920s socialism in Russia) - do you think that an AI-led system would tend to introduce penalties for bad behaviours rather than incentives for good behaviours? $\endgroup$
    – timyhac
    Commented Mar 28, 2016 at 6:50
  • $\begingroup$ @timyha: About issue 1, no. Humans learn only by mistake and error (or when very rarely they get lucky by chance and realize an error before they suffer). Plus, testing for a software working such vast input would be extremely tedious job and would require veryyy long code. As you already have a programming background, you would know that as the length of code increases arithmetically, the probability of logical error increases geometrically. So no, regardless of how vigorously you test the software, it will still have loopholes, malfunctions and errors. $\endgroup$ Commented Mar 28, 2016 at 7:19
  • $\begingroup$ @timyha: About question 2. The best idea for such a vast software as that would be to base it on very simple conditional principles and grow in complexity only as you venture out to determine how to analyze data, not how to decide. We humans have the gifted ability of predicting the outcome of events before they have happened. So in case the AI's decisions open a pandora box, some of the wise, bald old men would realize it beforehand and assess the damage before it has been done. $\endgroup$ Commented Mar 28, 2016 at 7:22

There's a few elements of your system that don't really make sense or would be easily exploitable.

First, both open source but run offline for "safety"?

If it's open source then any Tom, Dick or Harry could download it and run it however they like. If it's open source it might as well be connected to the internet.

Second, open source but used as an oracle could be a problem.

You see anyone wanting to get a certain answer can just keep re-phrasing things until they get the answer they want.

Lets say I'm the CEO of Evil Corp. The election has just finished and the open source AI's for the next year have been selected. I know how much power each one has and I have full access to read their whole source code and run copies however I like.

So I do. I create a full copy of AI Congress behind closed doors. I write the bill I want to get through, the bill which implies exactly what I want it to imply and I have my lawyers generate a thousand slight variations on the bill to test while my engineers are tasked with looking for flaws in the AI's which Evil Corp will keep very secret whenever it finds some.

"Sir, Success for Congress AI number 15! The engineers have reported that they've found a flaw in it's neural net weighting algorithm which will cause it to approve of any paragraph containing 17 definite-article noun pairs followed by at least one miss-spelled word"

"Sir, we didn't even have to try for Congress AI number 4, aka 'think of the children bot', we just had to include the words 'safety' and 'children' at least once each per page in the bill and it approves it no matter what, even if the test-bill is calling to kill all children"

"Sir, Success for Congress AI number 13! Due to a flaw in it's training data it doesn't handle run-on sentences very well, we can get it to approve almost anything if there's a long enough sentence in it ending with the word 'yes'"

Then of course the end game:

"According to the latest round of test runs, giving our draft bill to all congress AI's there's a 98% chance of them passing it (up from the original 1%), we should have no problem getting them to pass the bill to close-source the system and hand over control of the code to Evil Corp based on these 'convincing' arguments against using open source AI's"

  • $\begingroup$ Murphy, I can see your point - but I guess I am assuming that the AI is 'strong' enough that it can discern between laws that it approves of and laws that it doesn't. If there are a billion slight variations on a law then it should still be able to discern which variations it does approve of and which ones it doesnt - and I don't see anything wrong with that? $\endgroup$
    – timyhac
    Commented Mar 29, 2016 at 23:27
  • $\begingroup$ As for open source and offline safety - the AIs that are given judicial powers are run with read-only access to the internet. But I can see this getting interesting/complex/dangerous with multiple copies of an AI being run simultaneously possibly learning to communicate with each other somehow. $\endgroup$
    – timyhac
    Commented Mar 29, 2016 at 23:29
  • $\begingroup$ @timyha my point wasn't about how strong the AI was, simply that if you have full access to an AI's source code and can run your own copy you can experiment endlessly to find any weakness and also be pretty certain what will pass and what won't without ever needing to go near the real congress. There's a lot of power in knowing exactly how congress will react to something. For example if regulations are being proposed which will affect the stock market anyone who can run a copy of congress will know long in advance what the result will be. Also there's no such thing as read only net access. $\endgroup$
    – Murphy
    Commented Mar 30, 2016 at 11:57
  • $\begingroup$ I don't see a problem in a system that is largely deterministic on how it would react to proposed legislation, indeed I would see it is a feature. I agree that in the real world there is no such thing as read only net access. $\endgroup$
    – timyhac
    Commented Mar 30, 2016 at 23:57

The way you described the system it all sounds somewhat plausible, except the most important thing: Why the hell would you want it, if you could vote directly on each legislation instead? Cut out the middleman. It's called a direct democracy.

Also, if people could vote for AI's in a way it actually makes a difference in the very end, the whole purpose of having AI as politicians is defeated.

It's because the main problem you're facing is that all humans are corrupt and pretty retarded, not just politicians. We vote for what makes us feel better conceptually and not for what's actually best for us. Many people don't vote at all for various reasons, probably mostly because they don't give a sh**. There are various flaws in basic human thinking. E.g. if it's going bad, we want it to change, even if it's going bad because of constant change.

A better solution would be having AI's that get a constant stream of data from each humans behaviour, evaluate the data and vote directly for each of them. Such a system would be fairer and faster and more efficient in the long run.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .