Tip: click a paragraph to jump to the exact moment in the video.

1. 00:02 Autism spectrum disorders are little understood phenomena. Myths abound, misinformation flourishes. And today I would like to discuss autism spectrum disorder as a metaphor. For example, in the study of artificial intelligence and whether autism spectrum disorder is properly
2. 00:27 understood when people outside the profession are using it. One very common belief about autism is that it involves problems in communication. Autistic people are either miscommunicative or non-communicative. They fail to interact with what is colloquially known as
3. 00:54 neurotypicals and they get social cues wrong. They misread other people, other people's body language and other other people's proper language. This is the belief at least. Today we are also going to discuss a study which seems to refute this um attribution seems to challenge
4. 01:19 the idea of the belief that autistic people cannot communicate somehow. So putting the two together, we are going to study the transforming landscape of our understanding of autism spectrum disorder. My name is Sam Vaknney and I'm the author of malignant self- loveve
5. 01:41 narcissism revisited. I'm also a professor of psychology. Last month an article was published in in a in a arxiv arceive and the article is titled neural howl round a large enlarge language models a self-reinforcing bias phenomenon and a dynamic attenuation
6. 02:08 solution the author is Seth Drake and this is how neural howl round is defined mind. New Harl says the author, more formally described as recursive internal salience misreinforcement, RISM, is a failure mode arising specifically and directly from
7. 02:33 self-reinforcing probability shifts within an LLM large language model based agents internal state. Now this sounds very very discombobulating and off-putting but trust me it's an amazing concept amazing discovery actually and later on the author tries to align it
8. 02:55 with our current understanding of autism spectrum disorder which I will challenge in the second half of the video. So the author identifies four key characteristics, four key components unique to the phenomenon that he descri he he he labeled neural howl
9. 03:16 round. The first characteristic closed feedback loop. Unlike model collapse, says the author, which results from generational degradation across training cycles, neural howl round could emerge within a single model instance during real time inference, which is pretty
10. 03:40 terrifying. The second component, the second element is salience waiting trap. Whereas confirmation bias reflects biases inherent in training data says the author. Neural how round does not require bias training data to occur. It can develop spontaneously due to
11. 04:00 internal reinforcement dynamics within inference itself. Number three, cognitive rigidity. While bias salience s weighing often results from data set skew wrong data, neural hull run may arise even in perfectly balanced data sets if specific responses become dynamically reinforced
12. 04:27 mirroring similar effects observed in human cognition. And finally there is in this neural how round a new phenomenon in artificial intelligence there is a self-perpetuating distortion. Neural hard round represents an intrinsic distortion of salience waiting that
13. 04:50 recursively perpetuates itself once it reaches a critical threshold leading the model further into lockedin state of false overconfidence and response fixation. And if this sounds like some people you know, it's because these are also common phenomena in human cognition
14. 05:10 among people. And the author amazingly compares it to autism, autism spectrum disorders. He calls it digital autism. I think a much more appropriate label would have been artificial intelligence autism or AI autism. Here is what the author has to
15. 05:30 say. We recognize and acknowledge the sensitive nature of comparing neural how round to autism spectrum disorder in humans ASD. We do not suggest that neural how round constitutes autism in any biological or human sense. However, we believe that it exhibits a functional
16. 05:50 analog, a pattern of information processing that mirrors certain cognitive traits associated with autism spectrum disorder. Specifically, we propose that an agent experiencing neural howl round may exhibit behaviors that to an external perspective may
17. 06:10 resemble traits often associated with autism spectrum disorder. This was the inevitable and mandatory political correct statement. And now to the point, the author compares this glitch or this bug in artificial intelligence to autism spectrum disorder. He says that they
18. 06:29 share certain features. Number one, fixation. Autistic individuals often experience an intense focus on specific topics or interests and this focus appears unshakable. Similarly, an artificial intelligence agent experiencing neural halm will experience constant
19. 06:54 reinforcement of a small subset of responses. until they dominate all outputs forming a self- sustaining fixation. Second feature in autism context inflexibility. Some autistic individuals may struggle with shifting between multiple convent conversational topics
20. 07:16 often returning to preferred subjects. Likewise, neural hull realm would distort salience prioritization, causing non-reinforced topics to be overlooked and not prompting a response from an AI agent, thereby reducing the agents ability to adapt and respond to diverse inputs. In
21. 07:41 other words, the AI agent remains fixated on specific responses and topics very similar to an autistic person. The author continues, "Cognitive overload. Many individuals with autism spectrum disorder experience sensory and cognitive overload in highly stimulating
22. 08:02 environments, and this impacts conversation and decision-m in an AI agent. Excessive recursive salience signals may overwhelm and drain processing resources causing abbreviated and possibly simplistic responses and reduced nuance. It might also result in an increased risk of
23. 08:27 hallucination. The next element in in this new discovery regarding artificial intelligence resembles autism spectrum disorder as well. Perseverative thinking. Perseverative thinking, sometimes called hyperreflection, is a state observed in some individuals with autism spectrum
24. 08:50 disorder, where a cognitive process continues indefinitely due to an intrinsic intrinsic perceived need to refine it to perfection before proceeding. It is possible that an AI agent could become trapped in an unbounded recursion of infinite refinement loop and in this state
25. 09:13 response finalization would be perpetually deferred leading to an apparent cognitive stall and perceived withdrawal. And finally, there's executive function loss. Autistic individuals may experience cognitive overload when faced with options which
26. 09:31 appear equally valid and this results in an inability to prioritize or act in a state of apparent withdrawal. Similarly, an AI agent could experience salience collapse and be unable to resolve an input as all probability weights are equalized. This
27. 09:52 would effectively lock the system into a permanent undifferentiated state. An artificial intelligence affected by salience collapse would become nonresponsive not due to a lack of available outputs but because not single output option emerges as a distinct resolution pathway.
28. 10:14 So this was a comparison between autism spectrum disorders or disorder and um this glitch or bug in artificial intelligence which has been newly christened um the neural um the neural um problem that has been described in the article and the author compares the
29. 10:39 two. He compares neural hull round to these various deficiencies and defects and dysfunctions in autism spectrum disorder. But are these dysfunction real dysfunctions real? Or are they just beliefs that we have? Some ways to illustrate and visualize what's
30. 11:00 happening in autism spectrum disorder because the disorder is so opaque, so impenetrable. Is it just a metaphor? All the all these dysfunctions, are they just metaphors or maybe just speculations? One of the most common beliefs about autism spectrum disorder
31. 11:19 is that people with autism, autists, are incapable of proper communication. They have severe communication problems. But this is an example of a belief about autism spectrum disorder that has never been thoroughly investigated similar to all other attributes of autism spectrum
32. 11:42 disorder. Recently an article has been published and the article is titled information transfer within and between autistic and non-aututistic people. The authors are Crompton, Foster, Wilks and others. and it was published in nature human behavior 2025 this year. The
33. 12:05 authors claim that there is no significant difference in the effectiveness of how autistic and non-aututistic people communicate. They challenge the stereotype that autistic people struggle to connect with others. The findings in the study suggest that
34. 12:23 social difficulties often faced by autistic people are more about differences in how autistic and non-aututistic people communicate rather than some inherent or innate lack of social ability in autistic individuals. So it's about the interaction between autistic and
35. 12:44 non-aututistic people. It's not a problem in autistic people. And so there is a stigma around n autism of course and there are no
36. 12:59 effective there's no effective support for the development of communication skills across the bridge between autistic and non-aututistic people or as they're non-coloqually neurode divergent and neurotypicals. I hate these labels. Autism is a lifelong disability in many
37. 13:21 ways. There is there is a divergence from from a typical brain. It's a neurodedevelopmental disorder. In this sense, it's it's a medical problem, not really a psychological one. People exper people with autism experience the world and interact with it in ways which are
38. 13:40 alien to us. Autistic people often communicate more directly. for example, they're blunt. They may struggle with reading social cues and body language as I mentioned. And so this leads to differences in how they engage in conversation compared to non-aututistic
39. 13:57 people. The experts who wrote this article all hailed from University of Edinburgh. Um they studied 311 autistic and non-aututistic people. These were divided into groups and some groups were purely autistic, others were purely non-aututistic and some groups
40. 14:18 were mixed and the design of the study is is pretty fascinating. The first person in the group heard a sto heard a story from the researcher. The researcher told the story to the first person in the group and then this first person passed it along to the next
41. 14:35 person. Kind of a broken telephone. Each person had to remember and repeat the story and the last person in the chain recalled the story aloud. And so this allowed this permitted the scholars and researchers and experimenters allowed them to measure to
42. 14:55 quantify the amount of information that has survived authentically and genuinely and faithfully across the chain. The amount of information passed pass passed on at each point in the chain was scored and this gave an indication of how effective participants
43. 15:15 were at sharing the story. In other words, at communicating and they found the researchers found that there were no differences, shockingly no differences between autistic, non-aututistic and mixed groups. The information survived to the same level of
44. 15:34 faithfulness across all these groups. In other words, autistic people did not have any difficulty to comm community communicate information and to preserve it across large groups of people. After the task, the participant participants were asked to rate how much
45. 15:53 they enjoyed the interaction with the other participants. They were asked whether other participants were friendly or easy or awkward and whether the exchange itself felt you know bad or and so researchers found that non-aututistic people prefer to interact with others
46. 16:11 like themselves. Non-aututistic people preferred non-aututistic people and autistic people preferred learning from fellow autistic individuals. In other words, birds of feather flock together. Autistic people found it much more easy, much easier and much more expedient and
47. 16:30 and and and much more um efficient to communicate with other autistic people, whereas non-aututistic people had the same feeling when interacting with other non-aututistic people. And this is likely because the the autistic people and non-aututistic people um communicate
48. 16:55 differently. It's not the quality of communication. It's not the quantity of communication. It's not the faithfulness faithfulness of communication. It's not the efficacy of communication. It's that autistic people and non-aututistic people have different ways of
49. 17:12 communicating, different modalities. Um the same researchers by the way conducted a smaller study earlier on and the results confformed the new evidence seem to corroborate the earlier study. Dr. Katherine Crompton uh one of the lead authors um said that autism has
50. 17:38 often I'm quoting autism has often been associated with social impairments both colloqually and in clinical criteria. Researchers have spent a lot of time trying to fix autistic communication. But this study shows that despite autistic and non-aututistic people
51. 17:56 communicating differently, it is just as successful. With opportunities for autistic people, often limited by misconceptions and misunderstandings, this new research could lead the way to bridging the communication gap and create more inclusive spaces for all.
52. 18:14 Let me read to you the abstract of the bigger of the larger study. Autism says the abstract is clinically defined by social communication deficits, suggesting that autistic people may be less effective at sharing information, particularly with one another. However, recent research
53. 18:33 indicates that neurotype mismatches rather than autism itself degrade information sharing. And so this is what the study has substantiated. The study proved this hypothesis that actually the problem is in communication between different types of people autistic and nautist
54. 18:58 non-aututistic. This abstract says we found no difference in information transfer between single neurotype and mixed neurotype chains. Non-aututistic chains indicated higher rapport and disclosing diagnosis improved rapport. This result challenges assumptions about
55. 19:16 autistic communication deficits but contrasts with prior findings. Enhanced participant heterogeneity and methodological differences may explain these unexpected results. And so it seems that the conclusion is that people with autism spectrum disorder
56. 19:37 communicate perfectly. They also communicate perfectly with non-aututistic people. However, they communicate more efficiently and they prefer to communicate with other autistic people as do non-aututistic people who are perfectly capable of communicating with auti autistic people
57. 19:57 but prefer not to. There are silos of communication and autistic silo and nonautistic silo. So the study was published on May 14th in nature human behavior and let us loop back. Let us go back to the artificial intelligence study. The author of the art of the
58. 20:20 artificial intelligence study, the one who came up with the with the discovery of the neural howl round. Yeah. the the glitch or the bug in artificial intelligence agents that causes them to fixate and being unable to entertain a diversity of responses and views. So the
59. 20:39 comparison that the author Seth Drake has made to autism relies on these beliefs on these perceptions on these stereotypes. And if these were to to be challenged meaningfully, some of the comparison may be spurious, maybe wrong. It seems that cognitive processing and language
60. 21:04 processing in autistic people are not as alien as we seem to believe. And whatever is happening in artificial intelligence may resemble outwardly and perhaps functionally autism spectrum disorder but does not capture the internal dynamics in autism. And in some cases the imputed
61. 21:31 assumptions are completely wrong. For example, the assumption that autistic people cannot communicate properly that they have a problem with communication. Now we know that they don't have a problem with communication. They just prefer to communicate with other
62. 21:47 autistic people and not with non-aututistic people. This preference of course affects or influences the efficacy of communic of the communication, the motivation to communicate, the the passion in communicating, the the effort put into communicating specific data or
63. 22:09 information. And when you don't want to talk to someone, when you don't want to communicate to someone, when you'd rather communicate with others, of course, the communication suffers to some extent. And it seems to be um uh the core issue the the the
64. 22:28 preference of autistic people to communicate mostly and only with autistic people and how they have a preference to not communicate with non-aututistic people. This could be extended to artificial intelligence. By the way, artificial intelligence has been proven to be
65. 22:49 suggestible and artificial intelligence definitely agents definitely develop preferences for specific users or specific classes of users. In this sense, perhaps we're coming back full circle and it's really a kind of autistic mindset that artificial
66. 23:09 intelligence has. In the description, you will find a link to a um playlist of
67. 23:19 other videos on artificial intelligence. Follow the link and listen to what I have to say on this emerging amazing and very threatening technology. A technology that abinitio from the very inception emulates and imitates human beings, human mind and the human brain.
68. 23:42 An act of simulacum and mimicry that could culminate in a replicant, in a clone, in an android, in a copy of a human being, a simulation that would be good enough to deceive real human beings. And this raises issue with the uncanny valley reaction and so on so forth. I
69. 24:04 will not go into all of this but it raises fascinating topics. To remind you the diagnostic criteria for autism spectrum disorder include lifelong impairments in social communication and interaction that contribute to broad social disabilities and poor functional
70. 24:25 outcomes. Number two, these difficulties are associated with fewer friendships. Number three, less social support. Number four, loneliness. Number five, challenges securing and maintaining employment. Number six, seven, and eight, poor mental health. And number
71. 24:43 11, reduced equality. The majority of research so far assumed a deficit model of autism characterizing the differences in autistic socialability and communication as deviations from normality in need of fixing in need of remediation. But this model, this deficit model ignores the
72. 25:06 relational nature of social interaction. In other words, it's not only the autistic person who is to blame. It's not only the problem lies not only with the autistic person but with the people with whom the autistic person is interacting. The deficit model locates
73. 25:24 the cause of social interaction difficulties exclusively within the autistic person which is never right. There's a growing body of research and that shows that the factors that influence how autistic people interact with non-aututistic people and um are v
74. 25:44 varated. They're varied and this has an impact on autistic social experiences. Communication of course is always birectional. It takes two to tango and it takes two to communicate as a minimum. Social difficulties experienced by autistic people can be exacerbated by
75. 26:03 the behaviors, reactions, social judgments and misunderstandings of non-aututistic people, social partners, intimate partners, friends, um others, colleagues and so on. So the stigma attendant upon autism, stigma attached to autism is a kind of projection. It's
76. 26:26 it's as if non-aututistic people exonerate themselves. They say we are not the problem. We are perfect. The autistic people is to blame. The autistic person is defective. The the problem is with the autistic person. I as a neurotypical, I as a non-aututistic
77. 26:46 person, I am perfect. I am the standard. So non-aututistic people have been shown to form rapid negative judgments about autistic people on the basis of non-normative social pres presentation. The autistic person presents himself or herself in a
78. 27:07 non-normative way and this engenders prejudice and bias stereotyping and so on. And these non-normative social presentations are strongly associated with a reluctance to interact with autistic people. Autistic people are shunned and ostracized from the get-go,
79. 27:27 which sometimes gives rise to compensatory postures such as narcissism. And so these judgments by non-aututistic people or the non-aututistic environment, these judgments reduce the social opportunities afforded to autistic people. They simply don't don't have
80. 27:46 with whom to interact and the frequency of interactions declines. Their barriers to achieving professional goals, social goals, personal goals. Non-aututistic people also miscon misconstrue and misunderstand what autistic people are thinking and feeling. It's all
81. 28:05 speculation. It's all fantasy. It's all imagination. when non-aututistic people and that includes scholars and psychologists and clinicians and psychiatrists and what have you when people who are non-aututistic including as I said professionals practitioners
82. 28:21 who are nonautistic are trying to imagine the internal world the the psychonamics of the autistic person they usually speculate they imagine it's a belief system it's not it's not scientific and so countless studies have documented the difficulties of
83. 28:40 autistic people. People when they are trying to access and when they attempt to interpret the mental states and emotions of non-aututistic people and of course the other side, the other direction when non-aututistic people have difficulties to access and to
84. 29:01 interpret the mental states and emotions of autistic people. This is not documented. There are no studies of this. Almost non-aututistic people struggle for example to identify autistic facial expressions. They can't decode autistic mental states. There's huge mis there
85. 29:23 huge there's a huge number of misunderstandings when autistic and non-aututistic people interact. Non-aututistic people dislike autistic people. They overestimate how egocentric autistic people are. Non-aututistic people even report being helpful to
86. 29:42 autistic people in circumstances when objectively no evidence of helpful behavior is demonstrated. In other words, they imagine they fantasize that they're nice to autistic people, that they are kind to autistic people, compassionate, helpful, and attentive
87. 29:56 when they're not actually. So given all this mess, given this context, autistic people feel more comfortable and relaxed in the company of other autistic people because they're rejected by non-aututistic people. Autistic people are more likely to want to spend time
88. 30:13 with other autistic people. They disclose more to autistic interaction partners and they empathize more with other autistic individuals. Autistic people also often communicate in a in a different way. They're much more blunt. They're much more direct. They're
89. 30:28 sometimes abrasive, literal, frank, and this is often taken badly. It's often misinterpreted as rude or even sadistic by non-aututistic people. But other autistic people welcome it. This is a preferred communication style by other autistic people. There are recent
90. 30:49 studies, they provided support for these conclusions. Autistic interactions feature non-normative social behaviors and these result in enhanced communication with other autistic people. Autistic people also use common social cues differently when they
91. 31:06 communicate with autistic or with non-aututistic people. The positive rapport produced between autistic peers extends beyond self-report can be detected even by observers. And so all these findings challenge the assumptions that impaired social communication and connection are
92. 31:27 inherent somehow innate in autism. These findings suggest that social and communication difficulties of autistic people arise probably from a mismatch between autistic and non-aututistic modes of communication, presentation and understanding.
93. 31:47 There are many empirical studies of the kinematic dissimilarity hypothesis of social interaction. It's a hypothesis that suggests that kinematic similarity is important for action predication and social interaction. They're well doumented kinematic differences between autistic
94. 32:06 and non-aututistic people. Recent studies found that autistic observers are more able to accurately predict autistic actions than non-aututistic actions. And conversely, non-aututistic observers are more able to accurately predict non-aututistic actions than
95. 32:24 autistic actions. It's like two languages and there's no dictionary. The data suggests that difficulties in social interactions arise because of objective differences in the way that autistic and non-aututistic people communicate rather than autistic social deficit. Say the
96. 32:46 authors the authors of this study um open the gates to new thinking. It's still early early stages. It's still the whole thing is in its infancy. The idea is to transition from a deficit model to a relational model to study to rigorously examine autistic to autistic
97. 33:13 communication and autistic to non-aututistic communication. And this could provide um a measure of or evidence of intact communication efficacy between autistic partners with selective breakdowns in communication occurring between autistic and non-aututistic people.
98. 33:36 There's a we could use for example a cultural learning paradigm to capture the transmission of information between autistic autistic peers, non-aututistic control pairs and mixed autistic and non-aututistic pairs. Going back to the study, Croton
99. 33:54 found and her allies found that autistic partnerships and I'm quoting facilitate interaction for autistic people. Specifically, one, autistic people transfer information to and from other autistic people as effectively as non-aututistic people do with one
100. 34:10 another. Two, the quality of information sharing selectively breaks down when one person is autistic and the other is not. And three, interpersonal rapport is higher within than between diagnostic groups. And these feelings accompany information information sharing benefits. And
101. 34:33 so there's an awakening to the fact that autism has a pronounced interpersonal dimension that interpersonal reactivity and interactivity, interpersonal relations, interpersonal communication are very crucial. that in many ways many of the symptoms of autism that used to
102. 34:57 be attributed to some neurodedevelopmental deficit are actually the outcome of deficient social interactions when non-aututistic people are non receptive to the idiosyncrasies of autism. And so we are entering a new territory, a new territory where we are beginning to consider
103. 35:24 autism in a new light. There is some fundamental template of neurodedevelopmental disability. That much is true. But it seems that the autistic person can compensate for it when this person is with other autistic people. So the deficit itself, the bug, the
104. 35:49 glitch, the disability, the neurodedevelopmental problem or challenge seem to be surmountable even when the environment is much more accommodating, compromising, receptive, understanding. And the problem is social rejection. And so the current study, the study I mentioned I'm
105. 36:20 analyzing also went further in a very important way. They explored whether being informed of the diagnostic status of another person's interaction another person in the interaction. If you inform the non-aututistic person that he he or she is communicating with
106. 36:38 an autistic person, would that make a difference? Knowing the diagnosis of the autistic person, does it affect and influence the way the non-aututistic person communicates and relates to the autistic person? Does it impact information transfer? And so participants were
107. 36:59 informed of the diagnostic status of their partner before the interaction. In real world interaction, many autistic adults do not disclose a diagnosis because they're afraid of bias and discrimination and worse. And so there's emerging evidence that suggests that
108. 37:16 awareness of an autism diagnosis influences how autistic people are perceived by non-aututistic people. and but it doesn't influence how autistic people are perceived by autistic people. In other words, when you communicate your your diagnosis of autism to a
109. 37:35 non-aututistic person, it's likely to have dire consequences. When you communicate your diagnosis to an autistic person, it wouldn't make a difference. So, this these are terrible news. It's unbelievable that in the first quarter of the 21st century, we still
110. 37:57 relate to autistic people as paras or lepers or something. The efficacy of information transferred and the and the the the efficacy of the interaction I would say or sometimes of a whole relationship is impacted by some abstract label slapped on the on the autistic person.
111. 38:21 because there are these stereotypes and and so on and even in the artificial intelligence study with which I've opened this video there are stereotypes there there are nonsubstantiated beliefs speculations about the internal state of the of the autistic person
112. 38:39 uh autistic people have greater interest in and facility with information about factual systems. So this translates to efficacy of communication and rapport. When they discuss for example science or history or something they have the they have
113. 38:58 difficulty with communicating emotions and mental states. But the reverse the overse is true when it comes to nonautistic people. They're much less efficient in communicating facts and a little more efficient in effective communication. We need to accept this human diversity,
114. 39:18 not pathize it. We need to accommodate each other because these differences enrich the human experience and guarantee the survival of the species.