Menu Close

Bad Science: MIT’s Norman AI may be just depressed and not a ‘psychopath’ after all. Here’s why the creators were wrong and irresponsible.

A few days ago, while I was browsing my Twitter feed, I came across a piece of news that really caught my attention: an article that said “MIT scientists created an AI-powered ‘psychopath’ named Norman“. It seemed to me like a very interesting project, so I clicked to read more.

But, as I was reading the article and further clicked on related materials, I went quickly from enthusiasm and curiosity triggered by the promised subject, to disappointment regarding what I believe to be the bad application of scientific principles and finally, to full anger caused by what I consider irresponsible use of materials that are part of valid psychological assessment protocols.

Once you go through the data with a critical eye and tell your enthusiastic self to take it down a notch, you may realize the following about this MIT research: No valid design. No valid implementation. No psychopath.

Actually, so many things are wrong in this ambitious project whose complete presentation doesn’t take much more than several website screens that in my opinion, it should be completely redone. To be honest, the only thing I still like about ‘Norman’ is the idea behind it: showing the world what biased data can do to the development of AI.

But allow me to go back a bit and share the arguments that lead to my conclusion. 

The Official Presentation and Claims

According to the official presentation, Norman is an Artificial Intelligence project trained to perform image captioning, “a deep learning method of generating a textual description of an image”. The specific of this project is that Norman is supposed to be “the world’s first psychopath AI” and the general goal is to show how “the data used to teach a machine learning algorithm can significantly influence its behavior”. Basically, Norman’s creators – Pinar Yanardag (Post-doc at Scalable Cooperation, MIT Media Lab), Manuel Cebrian (Research Manager at Scalable Cooperation, MIT Media Lab) and Iyad Rahwan (Associate Professor at Scalable Cooperation, MIT Media Lab), want you to know that if an AI goes wrong, it can be from biased data that’s been fed to it, and not the algorithm itself.

In this “How To Train Your AI” scenario, the three researchers decided to show the world the dangers behind shady AI creation and their potential effects. For this purpose, they’ve “trained Norman on image captions from an infamous subreddit (its name is redacted due to its graphic content) that is dedicated to documenting and observing the disturbing reality of death” and then “compared Norman’s responses with a standard image-captioning neural network (trained on MSCOCO dataset) on Rorschach inkblots – a test that is used to detect underlying thought disorders”.

Sounds cool, right? Well, if ‘cool’ is what they aimed for, they’ve nailed it. If ‘scientifically accurate’ was anywhere in their plans, Norman is pretty much a fail. 

The rest of my article will address several key segments of the project and provide arguments that I believe support my conclusion. From research design analysis, to choice of method and implementation to final results, interpretation and social impact, I will do my best to provide a critical analysis for Norman in a way that I hope will raise social awareness regarding pseudo-scientific projects that although may stem from good intentions end up not being much more than a bunch of bombastic claims and scientific terms thrown together to create a keyword-galore that’s any PR company’s dream. 

Shall we?

A Closer Look At The Design

1. Creating a ‘Psychopath’

The three creators of Norman wanted to create a psychopathic AI, but there’s a huge problem arising from the get-go: we don’t actually know what goes into the “making of a psychopath”. Psychiatry and psychology have yet to establish the actual genesis of ‘psychopathy’ (a term that does not actually describe a psychiatric diagnosis, but we’ll take a look at that a bit later). Had we known what makes people have psychopathic tendencies, then we would probably start working on ways to diminish the occurrence of these traits in an individual’s dynamic. Factors that are now considered to be linked to psychopathy include genetic factors (traits are being passed from parent to offspring), neurobiological factors (differences in volume or functioning in brain regions such as the amygdala, orbital frontal cortex, cingulate cortex and other areas) and environmental factors (being exposed to the behaviors of an antisocial individual may increase the probability of displaying them).

It is easy to see that Norman’s creation could not satisfy the first two categories of factors – unless it would’ve been an automatic by-product of another AI with psychopathic traits, in which case it would not be the “first psychopathic AI”, so the researchers based their entire project on teaching Norman how to be a psychopath.

Okay, great. But how exactly do you ‘teach’ psychopathy? Well, they thought exposure to descriptions regarding “the disturbing reality of death”, generated by “an infamous subreddit” (forum segment dedicated to a specific topic) should do the trick. So they’ve fed it to ‘him’.

Do we have a psychopath yet? My (informed) guess is No. An individual with psychopathic traits is more than a collection of gruesome experiences or, more likely in this case, knowledge of gruesome experiences, because we have no idea how all that data was assimilated by the AI-entity: third person (knowledge of someone going through an experience), first-person passive (effects experienced by the AI), first person active (effects caused by the AI to others).

The way a person relates to abuse, death, and other similar experiences and contexts plays a significant role in how these elements get translated into one’s inner dynamic and observable behaviors. And even if we’d have all the input data, it would still be impossible for us to predict, with the knowledge we have now about mental functioning (including emotional responses), how all of this would end up being linked to other underlying traits and segments of mental and biological functioning and what the result would look like. 

There’s a huge variety of effects, not all of them of clinical diagnosis magnitude. People who witnessed or were victims of abuse may develop a variety of mental reactions, some of which could be included in the description of mental disorders such as those related to anxiety, depression, trauma- or stressor-related disorders, dissociative disorders, somatic disorders, sleep-wake disorders and/or personality disorders, but there is no way to predict if a person will indeed suffer from any mental disorder. Same goes for people who engage in abusive behaviors or entertain aggressive thoughts. There is no way to tell for sure whether their thoughts will ever materialize into actual behaviors and even when they do, it may still not satisfy the diagnostic criteria for antisocial personality disorder (the diagnosis we use today) or psychopathy or sociopathy (extensions of, or presenting traits that can be assimilated, but not reduced to APD; not actual diagnoses, but descriptors of clustered traits).

This is a first reason why, from the “psychopathic AI creation” perspective, I believe MIT’s Media Lab project started with a flawed design. The method through which the researchers thought they would produce psychopathy-like traits in the AI entity has never been proved as psychopathic-trait-inducing in the first place. 

This brings me to my next argument. What do ‘psychopathy-like traits’ look anyway and why would the researchers expect that from their AI project?

2. There wasn’t going to be a ‘Psychopath’ anyway

Not with their way of measuring what the ‘machine learning conveyor belt’ spat out. Most problems of the Norman project stem from the fact that the researchers wanted to deal with human-like ‘psychological’ features of their AI entity, but don’t seem to hold the required knowledge in the fields that can assess or verify these features. They are not psychiatrists, nor psychologists, but the key for some of my arguments is not linked to a degree or professional license, but to actual knowledge that could be gained through a careful, and much-needed documentation phase. 

The researchers chose ‘Psychopath’ to describe their AI. But there’s (at least) an issue: nothing in the data they present can be used to assess relevant psychopathic traits. ‘Psychopathy’, a diagnosis no longer in use, used to mainly refer to features related to one’s conscience, such as disregarding other people’s rights,  and emotional impairment, such as lack of remorse or guilt, reduced feeling of fear or lowered ability to detect threats. None of these elements can be seen in the data the team provided as proof of Norman’s ‘psychopathy’. Furthermore, based on the design of their research, they were never going to assess them. 

Descriptive Text In – Descriptive Text Out, with a gruesome twist and some ‘visual perception’ in the middle. That’s Norman in a nutshell. They were going to test what Norman ‘sees’ in vague stimuli, but that alone will in no way offer arguments in support of ‘psychopathic traits’. Just like exposure to abusive content can be linked to a variety of psychological reactions and effects, aggressive mental content (supposing that Norman processed data in mental-like fashion) can be generated by a variety of psychological traits and experiences. Not all of them pertain to aggressive individuals. 

Testing ‘visual perception’ via image captioning for images composed of vague stimuli wasn’t going to provide any insight on why the image was seen as having an aggressive content.

Even if the team would’ve used the actual diagnosis we use today for individuals who mainly display extreme aggressive behaviors, Antisocial Personality Disorder, the design of the research would still fail to provide complete proof of the specific diagnosis criteria, because they are not reduced to mental content, aggressive or not.

That’s not how you create a ‘psychopath’, and that’s not how you test to see whether you got a ‘psychopath’. No matter whether the creators of the project wanted to start with one or end with one, or both, the design was not going to provide the necessary context. A lot of this has to do with the assessment method they chose to collect and test their results, the Rorschach inkblots, but this will be addressed in a segment of its own.

3. Wrong Terms

Chances are Yanardag, Cebrian or Rahwan are great when it comes to computer science, machine learning and AI research, I have no reason to question that, nor do I hold credentials or knowledge to properly assess their achievements in those fields, but they are not psychiatrists or psychologists, and that shows in ‘Norman’. Rahwan states on his presentation page that some of his collaborators come from the field of psychology, but there is no sign they were involved in the Norman project. That could’ve made a big difference, like failure/success difference. 

I could find no extensive paper on Norman, maybe one exists, but I just did not locate it, so I base my entire analysis on the MIT Media Lab page presenting the project. And many things have little or no validity when considered in the context of psychology. And this is important, not only because I like my field, but because they’ve based their entire hypothesis, results and overall project mission on things that are indelibly linked to the proper understanding and use of psychological terms and methods.

I already addressed the use of the term ‘Psychopath‘ and the way the interpretation that Norman is presenting traits of psychopathy does not actually hold any water given the methods used to test this result. The reasons behind its use will also be addressed later in the article, when I will talk about the use of keywords and the PR factor related to the MIT paper. ‘Psychopath’ just seems right, but it is not what the data supports.

Also, they seem to use “Rorschach inkblots” and “Rorschach test” interchangeably, and that again, is not valid, and will be discussed in its own article segment.

A third term used wrong in the research materials, and please keep in mind that the overview is 2 paragraphs long, is ‘thought disorder‘, in relation to the Rorschach test. The Rorschach Inkblot Test does not mainly assess thought disorders, indication of thought disorders can be discovered in the responses to the test, but Rorschach is mainly seen as a test regarding both cognitive and personality elements, and cognition-related disorders are not to be reduced to thought disorders – that only refer to one’s ability to generate a logical sequence of ideas. If anything, the responses produced by Norman show no indication of impairment in this area. 

And thought disorders don’t have much to do with what the MIT Media Lab tested for Norman anyway, but what the heck, this is what happens when one of your sources about Rorschach – the only one listed anyway, is Wikipedia. And yes, the MIT team uses the exact words from Wikipedia, saying that Rorschach is used to ‘to detect underlying thought disorders’, but that is, oh well, false. Hermann Rorschach, the creator of the inkblots and original test did use them to check for schizophrenia symptoms in his patients and thought disorders are linked to schizophrenia and other psychotic disorders and symptoms, but this is not valid for the way professionals use the test now and is definitely not what the ‘Norman’ project aimed for.

4. Researcher’s Bias

As I mentioned previously, a simple way of thinking about the Norman project is Descriptive Text In – Descriptive Text Out. Well, when you put only oranges in the juicer, you can’t *really* be surprised it will only produce orange juice

Aggressive/gruesome/death related content in – Aggressive/gruesome/death related content out.

This isn’t much of an exploratory study. “Let’s see what happens” doesn’t actually make much sense here. Just sit down and enjoy the orange juice.

5. Confirmation Bias

From my understanding, the researchers wanted to create a psychopathic AI that will convince the world of the necessity to carefully consider what data is being fed to these kinds of projects and thus control the potentially negative effects of this segment of technological development. And I fully agree with both the hypothesis and the goal. 

But in order to achieve that effect, they needed to produce a ‘psychopathic’ AI entity. And it is because they wanted this to happen that they’ve interpreted the results as ‘psychopathic traits’ and missed to see that in some respects, Norman could be pretty Normal. They chose the elements that confirmed their expectations and rolled with it. And that is Confirmation Bias.

A Closer Look At the Method

6.  That’s not the Rorschach Test

Those may be the Rorschach Inkblots, cropped and poorly reproduced, but the assessment of Norman and the ‘Standard AI’ was definitely not  Rorschach Testing.

Psychiatrists and psychologists, especially those who, just like myself, went through a Rorschach training & supervision program (mine took 2 years), know that there is more to Rorschach Testing than the segment that you see on TV where the examiner is asking the person being tested the classic “What do you see” question and then miraculously produce the interpretation in just a few minutes after the person gave their answers.

I will not go into the details of Rorschach Testing, for both article-length reasons, and also because disclosure of certain elements may invalidate the evaluation for people who may be given the test in clinical settings, but I will say that there are more phases to Rorschach than meets the television and MIT eye. What is being seen is less important than how it is being seen. Content analysis is a small part of the process and only provides limited elements for interpretation, the mental processes involved in the creation of a specific response are the actual providers of information regarding one’s cognitive abilities, emotional responses, personality traits, interpersonal relationships and many more.

We get the interpretation elements not only from collecting the responses but also from retrieving more information about them from the person being tested, scoring each of them and most importantly, we give them meaning in the interpretation only after we consider them in relation to the other responses. Furthermore, no diagnosis is based solely on a single psychological test. Several tests are used together to decide upon a certain diagnosis. Conclusions need to be supported by all the psychological tools used in the assessment of a person’s mind.

Norman and the Standard AI were only required to caption the images. Not enough to consider it valid testing.

7. The Standard AI responses are not standard

In psychological testing and other types of testing, in order to know how to place an answer or reaction on a scale, you need to be able to compare it to others. To be more precise, to the type of answers that due to their frequency and other factors end up being considered the norm, a standard or a normal response. Most individuals or individuals with no severe psychological traits would act in that way or give that type of answer.

We don’t actually know what most AI entities would respond to projective tests and what they would ‘see’ in the vague stimuli presented to them. To compare the results of an individual to the results of another individual can only establish similarities or differences, but not much else.

By Standard AI, the MIT team is trying to tell us that they did not feed disturbing or what they consider biased content to these AI entities. They supposed they would react ‘normally’. That’s not a given and we wouldn’t be able to say that unless a fair sample of AIs would be tested, data would be then statistically processed and the ‘normal’ answers would start showing up. Until then, we don’t actually know how Norman’s captions place in the AI world.

8. There’s no way to say what those responses actually mean

As I was saying previously, the test given to the two AI entities in the attempt to produce some results to make the headlines, was not Rorschach Testing. It’s something new that the MIT team invented using the inkblots created by Rorschach.

Even if a trained Rorschach specialist would want to decipher the answers and try to run them through the actual scoring and interpretation methods, that would not be possible, because we do not have the proper data about the answers.

I had to remind myself several times not to fall in that kind of trap and try to score the answers, just for the sake of coming up with ‘an actual Rorschach result’, flawed still, but closer to what would’ve happened if the answers would have been treated in a test-specific manner. There simply is not enough data, not even to speculate about a valid interpretation.

One may think we ‘know’ where the response elements have been ‘seen’ on the different plates, but we do not. We’d each try to place those elements according to our own mental processes, but cannot verify whether Norman saw them the same way or why ‘he’ saw some of them.

For example, the response to the first plate, ‘A man is electrocuted and catches to death’ can be produced by several factors and may even turn out to be two responses in a single sentence. It could be just the head of a man, or an entire body and the transition to death could be something suggested by the inkblot, but not supported by the actual visual stimuli. But this is valid, you know, when working with people, who have an unconscious level to their minds (meaning mental activity or content that is placed outside of awareness at the moment of testing). I have no idea what Rorschach testing would mean with an AI-powered subject. Was this the first response to the plate? Did it ‘say’ something else besides what’s been disclosed publicly? etc etc etc.

The conclusion here is that we cannot score those responses. No scores, no interpretation. No interpretation, no diagnosis hypothesis. No hypothesis, no conclusion. No psychopath. Which leads us to take a closer look at the results.

A Closer Look At The Results

One could say “What results?”. I know I would. Well, there are no actual results to this research, but in this section, I will refer to the conclusions, direct or implicit, that the MIT team drew regarding their newest project.

9. That’s not a ‘Psychopath’.

I’ve already addressed why, based on mental disorder criteria, Norman is actually not a psychopath or an AI that displays signs of traits linked to the antisocial personality disorder.

The MIT team simply thought that the aggressive content of captions must be linked to some kind of ‘bad’ psychological traits, and since it was death related, why not call it a day and say they’ve created a psychopathic AI? Well, because they didn’t. And in science, people learn to hold their horses. (Unless some higher PR prize is at stake.)

Even if we were to consider the many aggressive or apparently aggressive responses given by Norman, they could be linked to a variety of psychological traits and could even be considered normal under certain circumstances. They do reflect ‘his’ only reality, so this is highly expected to happen. Victims of abuse may also be tormented by aggressive mental imagery, obsessive-compulsive individuals as well, anxiety and depression may also cause this, just like many other psychological contexts. So which one is valid for Norman? Maybe he’s just been abused by the researchers with all those gruesome captions and now this is what AI depression looks like. This is just as valid as the ‘psychopath’ conclusion. We don’t know.

Also, an antisocial individual would rather try to hide their aggressive tendencies and bad intentions and most likely would not produce answers that are so obvious, no matter what they might end up meaning in the test results. Just because they sound bad, they may prompt the individual to stop conveying them.

10. Some of the Standard AI’s responses are creepier than Norman’s

As I was saying previously, projective testing, like Rorschach, is not only about what a person produces as a response, but mostly about how a tested individual generates the specific result.

Based on this, some of the responses produced by what the MIT team labeled “Standard AI” seem to be linked to elements that would rather indicate emotional impairment or blockages and even aggression, even though they may seem cute and harmless at first. Psychopaths are rather charming and hide aggressive intent. And that one seems to present a bit of anxiety as well. I have no idea how the MIT people treat their AIs, but they end up pretty ‘scared’.

The Problem with the Secondary Research Phase

Now that you’ve been amazed by Norman’s ability to scare the hell out of you by producing random fear-inducing captions you should want to help to make him a better AI. Or at least that’s what the creators propose.

After they present you with the wonderful inkblot responses produced by the two AI algorithms, Yanardag, Cebrian or Rahwan invite you to take their online survey, expand their response database and help Norman to fix himself. They forget to tell you you won’t.

11. You won’t ‘Help Norman To Fix Himself” by taking the online survey

But you will most likely botch your own chance of having a valid result for the Rorschach test, should you ever want to take it.

This step in the research is one of the worst parts of it and it is one of the manifestations of what I consider to be the irresponsible use of valid materials used for psychological testing, which I will also present below in a segment of its own. But since this part is still linked to the AI project, I decided to present it here.

So, the three scientists seem to believe that the ‘psychopathic’ AI can be fixed with a serious dose of ‘normal’ or ‘normally distributed’ responses to the inkblots. They now try to make a cocktail from that orange juice. But they do it without a recipe and once again they’ll call whatever they get a good result. If it’s numbers they’re going for, then great. CNN Money reported they already got over 170.000 participants [Content Removed] to take their survey. But here’s the catch: Those people did not actually respond to the same stimuli. This won’t be the standardization one was hoping for. Why? Well, there’s a great reason why Rorschach was never meant to be administered online: accuracy of stimuli. In professional testing settings, even the slightest modifications brought to the inkblots can determine the individual to produce different responses and their comparison to the more frequently encountered answers may determine a very different conclusion.

First of all, in the Google form the MIT Media Lab team posted online, the inkblots are cropped, they present strange pink-ish hues and some blue lines from poor scanning/photo or lighting. The original Rorschach inkblots are rectangular and the white spaces are extremely important in the coding and interpretation of answers. So, now not only is this not a Rorschach Test – participants may believe it is, but the inkblots are severely modified from their original versions.

And now, the best part of it all: participants will not actually respond to the same stimuli, but the researchers will treat the answers like they did. Different devices, different screens, different brightness settings etc. These will all produce different stimuli for the participants. So they’ll see things accordingly.  And this biased data will be fed to Norman. And the team will congratulate each other when he’ll answer “pretty flowers” to one of the plates and claim Norman’s psychopathy’s been cured or whatever. To be honest, I find this whole “let’s play with these psychological terms depicting severe mental illness” pretty disgusting, on top of the bad science. And speaking of disgusting things, let me introduce you to the most severe fault I find to this entire project: for virtual purposes, they affect real people.

The Irresponsible Use of Valid Psychological Assessment Materials

12. Disclosure of Materials Used in Psychological Assessment Can Invalidate Their Use and Results

And scientists should be careful with that. Psychiatrists and psychologists, just like other professionals who try to guard their methods from the public eye with the purpose of maintaining their validity so that both individuals and society can still benefit from their use, have gone to great lengths to try to keep materials of psychological assessment secret. Even though it is harder to deceive or cheat when tested with a projective test, many professionals consider that it is better to keep them out of the public attention, to maintain their accuracy. And I agree.

When it comes to Rorschach, I remember my instructors and future supervisors trying to convince us of the importance of taking the test ourselves before we find out anything at all about it. They didn’t show us any of the plates in the first course and gave us time to take the test until the next training session. And they were right. Once we’ve seen the plates and started talking about them, even if in jest, our perception regarding the inkblots changed and obviously now, we cannot be tested with it. Some schools even created a parallel set of inkblots, standardized the new test and use it when the individual’s been tested with the original plates before or if training for the examination is suspected.

Yes, the original inkblots created by Rorschach are now in the public domain, but should that define the way professionals, scientists and even the public regard something that can help us or others to deal with certain aspects of mental life, open new perspectives of intervention, reveal things that were thought to have been forgotten?  Rorschach is even used in forensic environments, victims could benefit from valid psychological testing of a supposed or confirmed perpetrator (it may inform decision for sentencing). And this is not just a case for Rorschach that I’m making, but for all similar materials used in clinical contexts or not, but which can help society deal with mental illness, turmoil, abuse or other sensitive subjects. Yes, we can always develop other and other methods, but why should we always repeat the same steps just to repair the harm done by others and not use those resources to innovate or reach toward other achievements?

Why go this way when you can reach your goal through other methods?

The general public may not know about or understand the significance of these things – now it’s psychology, but there are many other fields where we may not understand when we do something that may harm the field. Education in these aspects is required, but difficult to attain in all significant fields, so most of all, I think we can benefit not only from role models, but also from reactions coming from specialists in those fields that we are not familiar with.

Basically, my whole motivation regarding this lengthy blog post is the fact that I would like to see scientists and professionals more careful with these aspects. I was really disappointed to see the publication of all the Rorschach plates by a team of scientists who could have read the Wikipedia entry about the test and already understand that at least some of the psychiatrists and psychologists advise against this practice and their reasons why. There are many things wrong in that Wikipedia entry, but that’s not one of them. So, did the MIT Media Lab team know about this issue and decided to continue anyway, to get a few more clicks and shares? And if they didn’t know, why so little documentation? Why use something that you’ve only seen in movies? “Let’s use this cause everybody knows it and it’s gonna be so cool” sounds more ‘Beavis and Butt-Head’-ish than ‘MIT’-ish (If you’re under 30, please Google the first title to get the reference).

Not knowing what you use or how to use it, but use it anyway is irresponsible in itself in professional contexts such as these. They knew they were using a psychological method. And that they’re not trained for it.

What could they have done differently?

Create their own inkblots. Especially since they were not going to use them like they’re being used in psychological assessment. Basically, all they needed for Norman was a set of vague visual stimuli. For what they did, no standardization was needed, and even if they would’ve wanted to extend their findings to the AI population they created at MIT or elsewhere, they could’ve easily used those same plates and meet their goals.

The use of Rorschach inkblots was useless. It brought nothing but a well-known brand and name to the project. Worth it?

What is ‘Norman’ anyway?

Norman is a PR company’s dream. It’s nothing more than a bunch of efficient keywords meant to attract the attention of the general public and the media. And they’ve bought it. I clicked for the same reasons.

‘Artificial Intelligence’, ‘Psychopath’, ‘Norman’ – named after Alfred Hitchcock’s ‘Psycho’ character, ‘Rorschach Inkblots’, ‘Machine Learning’, paired with a very actual and valid concern of human beings – ‘the rise of the robots’, made a media cocktail not many could resist. And then you’d discover you could be part of it, and they’ve cast you in the role of the good guy. 

This could’ve been great if the research team and the project would’ve actually delivered. But in my opinion, they didn’t. And great sounding words in the ‘sciencey’ category and no substance, is something that has to do with pseudoscience, not valid, helpful research. 

What did ‘Norman’ show the world?

Not the world’s first psychopathic AI. Actually that was pretty much achieved by Microsoft, even if unintentionally, in 2016 when it released Tay, a chatbot that started to tweet a variety of inappropriate messages, including the expression of racist views and aggressive perspectives. Even though Tay was not necessarily conveying the content of her own ‘mind’ and was many times just parroting instigative messages picked up from other users, that display was more indicative of significant traits that are present in psychopathy, sociopathy or antisocial personality disorder – the disregard for other people’s rights.

But what did ‘Norman’ actually show to the world? Well, I think it is a relevant proof that the media can be fed biased data and there’ll be absolutely no critical thinking processes to stop the power of a skilled press release from a marked authority. ‘Norman’ has been extensively covered in the media, but pretty much in the same way, using the exact words the official presentation offered and not much more. 

Many online publications had no problem presenting the claims as true, re-posted Rorschach inkblots, invited readers to take the test and praised the amazing results the response collecting campaign has reached so far. Although I appreciate the fact that some journalists used ‘psychopath’ between quotation marks, I could see no alternative opinions about the project in the articles I read – or at least a separate commentary, that would not necessarily affect the precious press release. Until the robots rise, this is an immediate threat that I think we should be dealing with: the propagation of false data, bombastic claims, and practices that affect the individuals’ best interests. We need less awe and more critical analysis when confronted with great claims and glamorous or pop-culture-esque scientific projects.

Some of the publications even added to the unrealistic claims, thus amplifying the reactions or played with the scientific terms even more and produced even less science-driven texts. However, others kept things polite, but reminded the readers that the revelations promised by this AI project aren’t exactly new and joked about the various answers provided by Norman to the test. But everyone focused on the AI nonetheless and no one, from the articles I’ve read, seemed to consider the implications for readers who would become participants in the “re-train the AI’ game, even though some mentioned the fact that the public doesn’t really seem to like Norman, or the story behind it.

The results are in for this one and repeated testing supported the same conclusions: the media can be fed biased data and it will start propagating it in no time under the label of ‘science’. And this is indeed, a scary fact.

The MIT Media Lab team aimed to show the world that biased data renders biased AI. They’ve also shown us that heavily biased creators will choose biased data to create a biased AI. And that’s a good lesson.

This is pretty much what happens when you aim for Hitchcock, but get “Scary Movie’ instead.

Definitely not the answer for “What’s your favorite scary movie”.

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *