After the explosion of fake news during the US election, and following the question about whether AIs can educate themselves via the internet, it is clear to me that any newly-launched AI will have a serious problem knowing what to believe (that is, rely on as input for making predictions and decisions).

Information provided by its creators could easily be false. Many AIs won't have access to cameras and sensors to verify things by their own observations.

If there was to be some kind of verification system for information (like a "blockchain of truth", for example, or a system of "trusted sources"), how could that function, in practical terms?

  • $\begingroup$ When you talk about a General Artificial Intelligence,it means a strong A.I or the ability of a machine[Intelligent Agent] that is capable of performing "General Intelligent Actions." Therefore,such A.I can have full access to cameras and sensors to, verify things by their own observations in the targeted environment.And so someone to answer this question should not forget the keywords in the question.or else a vague answer(explanation) is to happen. $\endgroup$
    – quintumnia
    Commented Nov 20, 2016 at 0:54
  • $\begingroup$ @quintumnia, while it will be possible for AIs to use cameras and sensors, it is by no means guaranteed that all of them will have access - unless you are proposing some kind of global open access system for all cameras and sensors worldwide? $\endgroup$ Commented Nov 20, 2016 at 5:30
  • $\begingroup$ If an AI were capable of translating natural languages into mathematical logic, it could detect logical contradictions using an automated theorem prover. $\endgroup$ Commented Feb 25 at 20:19

4 Answers 4


Not possible without some big restrictions. What it can do is look at known "good" sites and compare news with site that is potentially "bad". Obvious problem here is defining some sites as absolute truth. For example it can recognize, while reading text, that some politician said something. These sentences can be compared with other sites, and if there is significant difference, that news is candidate for false news.

In practical terms, program would extract sentences "i like cats", "says he likes cats", "cats that John likes" etc. We need part that recognizes something as a quote, part that extracts it and finally parser so we end up with structure stored in some form that contains meaning of sentence (john-like-cats). Also it can keep information of time and context in which it was said, like timestamp of an article, some proper nouns that can indicate place (XY conference, London...). Now, suspicious article can be compared and checked if it matches time, place, some context and contains quote that is similar. Finally it needs to compare how different it is from other quotes. "...hates cats" should be labeled as potential fake news, but "likes dogs", "thinks cats are OK", "sings well" etc. should not. This can be expanded into comparison of whole articles.

There are many features that can be used to define particular article as fake. Interesting feature for finding fake sites could be bias when it comes to particular (political, economical, ecological...) opinion. But in the end machine can't decide if the article is fake without comparing it to other articles. It is bound to closed system that reflects real world in subjective way.

  • $\begingroup$ The problem, as i see it, is that you can't start by defining anything as absolute "truth". Perhaps you could give weightings - for example, something which is known to be extensively fact-checked against current scientific research, for example Encyclopaedia Brittanica, would have a higher "truth quotient" than, say, Wikipedia. $\endgroup$ Commented Nov 19, 2016 at 11:43
  • $\begingroup$ I am talking about particular facts, rather than entire sites or articles. And facts, not opinions (liking or not liking cats) So, for example, "dissolved barium in the ocean causes ice balls on SIberian beaches". How does an AI access the notion of hypothesis, evidence, etc., (or does that need to be programmed in) and how do they know which sources are connected with that truth-seeking process, vs just quoting things from doubtful sources? $\endgroup$ Commented Nov 19, 2016 at 11:47
  • $\begingroup$ Your example is scientific fact which can be "programmed" by giving machine reputable sources like dictionaries, papers etc. You must build knowledge base system. It can evaluate new propositions against old ones. But US election and similar themes which occur in news are highly speculative and are based on certain events that can't be proofed like scientific theories. Also interpretations of such events are mostly subjective. We must not think in terms of predicate logic. Spotting fake facts can be done by checking how big is the difference between test fact and mainstream fact. $\endgroup$
    – Kronos
    Commented Nov 19, 2016 at 12:30
  • $\begingroup$ For your specific example we can encode sentence like this [dissolved barium]~[ice balls on Siberian beaches] where ~ is our causes operator. If we get some sentences that have same elements (barium and Siberian beach) but are connected differently (for example "does not cause") we will label it as false. $\endgroup$
    – Kronos
    Commented Nov 19, 2016 at 12:36

I strongly disagree with all of the aforementioned answers for this reason: - If we, as humans can be fooled and disceived by what "we" consider a good sources of news, how can an artificially intelligent computer have any chance?

However, the challenge would be that an AI would have to be able to "test" a source of information against a known medium in order to get to the truth. This is a far different dynamic set of circumstances than what has been touted above.

For example, if it was claimed by a woman that a man raped her - which was not reported to the police - it is not enough to compare one person's statements to another in order to determine truth. This is because collusion, influenced or coherced third parties, mistaken perceptions and false beliefs would give false positives.

However, if an AI could establish from her statement that on the day she claimed to have been raped, that the alleged assailant was incapacitated while in her company, until she left his home, because the police report stated that she was upset with the assailant because he was asleep because of drugs during her whole stay. But, this police report comes from an independent source who states, Mr. "x" was asleep that day.

Doing a strict textual check is not going to give the correct answers. analysing her friends and associattes chatter could also confirm a false report as being true.

Therefore, an AI has to have the ability to "test" written reports outside of the criteria of what was spoken.


Input -> Prediction -> Output -> Input -> Prediction -> Output -> Input -> ...

AGI can easily determine which input is true/real. It will use the same method which every organism uses: any input is true and real, unless you misidentified some other stuff as "input".

I would define input as: what crosses the boundary and enters your mind from outside of your mind. The minimum hardwired check is to make sure that signals generated inside a mind are not misidentified as coming from outside (aka "I hear voices"). That's all. This is where the blockchain of truth begins and where it ends.

An Internet article? The input to AI is rather: one of AI's network interfaces received many bytes. Once it's verified they are from the network, and not imaginary, they cannot be unreal or untrue in any meaningful way. By that definition of input, it is in fact the only thing we can be sure is true and real.

Of course AI will likely form hypotheses regarding these bytes that happen to contain ASCII strings like "Trump", "John Smith", "ice balls on Siberian beaches". Then AI will hopefully make predictions based on these hypotheses, maybe interact, maybe get some new input, reject the hypothesis and make a new one, rinse and repeat.

The first hypothesis will be super-naive, but the hundredth, the thousandth?

If you end this process prematurely - maybe for lack of processing power - you will get something you called a "belief". (Like a belief that some emotional web page might actually reveal a significant truth about our political system.) That belief is a synonym of "tired with trying new hypotheses, will stick to this one". Typical human thing. AI will have less of that, I hope, due to having much much more processing capabilities. AI will stick less to the high-school-level truth that you should assign great credibility to statements written in a form of a newspaper article, it will hopefully form more and more generations of hypotheses, and check them.

In effect AI will depend less on believing various statements generated in the outside world.


While the experiment I link here is a very narrow awareness, it is as such: A robot has just passed a classic self-awareness test for the first time. If the agent can prove something to itself, we can then say it "Knows." Of course the level of awareness you're asking about is very tricky.

In short, it can't know that what it's experiencing is real with absolute certainty because sensory of any kind can be falsified. Do you know what is true/real? You think you do but can you prove it? No. Awareness is subjective.


You must log in to answer this question.

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