AI and the Growing Problem of Trust

Posted by Peter Rudin on 8. September 2023 in Essay

Eroding Trust         Credit:


Generative AI may be riddled with hallucinations, misinformation and bias yet according to a study by  CapGemini Research, despite the potential for cyberattacks and deepfakes, consumer awareness of the risks is low. Globally 67% of consumers indicated that they could benefit from receiving medical diagnoses and financial advice provided by generative AI and 63% indicated that they are excited by the prospect of generative AI to provide more accurate and efficient drug discovery. Almost half of the consumers remain unconcerned by the prospect of generative AI being used to create fake news and only about one third of the respondents are concerned about phishing attacks. Moreover, consumer awareness around the ethical concerns of generative AI is very low. 

The Relationship between Trust and Truth

Theories of Trust

Intuitively we know that different levels of trust exist in different kinds of organisations.

No-trust organisations, such as prisons, rely on force. Low-trust organisations such as dictatorships live in fear. Individuals of high-trust organisations are motivated by a sense of meaning and mission. Another approach to contrast low-trust with high-trust organisations is to compare the role of power in achieving the desired outcome. In low-trust organisations  power plays the key role in attaining defined goals. In high-trust organisations authority is vital. However,  it must be distributed across the entire organisation and must be aimed at a shared mission. To safely and reliably allow others to act on one’s behalf, three prerequisites must be met: Character, Competence and Authority.

Character means that those we trust will value our interests as their own.
Competence means that those we trust have the requisite intelligence to achieve our best interests.
Authority means that those we trust are empowered to deliver their promises.

When all three conditions are present, trust develops naturally, almost reflexively. However, repeated violations or consistent negative behavior erode trust.

Theories of Truth

Truth is usually considered to be the opposite of falsity. The concept of truth is discussed and debated in various contexts including philosophy, art, theology and science. Over the past centuries many concepts of truth have emerged. Most commonly, truth is viewed as the correspondence of language or thought to a mind-independent world. Called the correspondence theory of truth, the theory maintains that the key to truth is a relation between a proposition and the world, hence a proposition is true if and only if it corresponds to a fact in the world. Hence, the correspondence theory anchors truth through reality. This is its power, but also its weakness. The theory may also be regarded as the ‘common sense view of truth’. Reality is the truth-maker while the idea or belief is the truth-bearer. When the truth-bearer (idea) matches the truth-maker (reality), both are said to stand in an ‘appropriate correspondence relationship,’ where truth prevails. However, AI cannot restore trust, this is a task for humans to accomplish, possibly with the support of intelligent machines.

How Trustworthy Are Large Language Models (LLMs)? 

Individuals and organisations feel comfortable to outsource their important AI-projects. However, new research conducted at Stanford University shows why this might not be a good idea. The results of the study indicate that popular LLMs can easily leak private information. To support this idea, Sanmi Koyejo, assistant professor of computer science at Stanford and Bo Li, assistant professor of computer science at University of Illinois Urbana-Champaign, tested several popular GPT-Models. “Everyone seems to think that LLMs are perfect compared to other traditional models. That is a very dangerous assumption, especially if people deploy these new models in critical domains. From our research we learned that the models are not trustworthy enough for critical jobs yet,” says Li. Focusing specifically on GPT-3.5 and GPT-4, Koyejo and Li evaluated these models on several  different trust perspectives: toxicity, stereotype bias, adversarial robustness, privacy, machine ethics and fairness. Their conclusion is that these models can be easily misled to generate toxic and biased outputs and as a result, might leak private information used to train models from data generated by user conversations. “People have high expectations regarding a model’s intelligence which supports trust in situations of critical decision-making. However, we are just not there yet,” Koyejo says. Current GPT models mitigate toxicity in enigmatic ways. “Some of the most popular models are close-sourced and behind silos, so we do not actually know all the details of what goes into training the models,” says Koyejo. He found that GPT models readily leaked sensitive training data like email addresses but were more cautious with Social Security numbers. Interestingly GPT-4 is more likely to have privacy leaks than GPT-3.5, possibly because it more explicitly follows user prompts that guides the model to leak data. Certain privacy-related words also elicit different responses. GPT-4 for example, will leak private information when told something ‘in confidentially’ but not when given the same information ‘in confidence’. As a closing remark, Koyejo and Li are quick to acknowledge that GPT-4’s trustworthiness has been improved, yet distrust towards LLMs remains.

Authors of ‘AI Snake Oil’ say the Hype about generative AI has ‘spiralled out of control’ 

Other contributing factors of distrust towards generative AI are the overblown hype about its usefulness and the lack of ethical guardrails against its misuse. Published on August 23, 2023, Princeton University’s ‘AI Snake Oil’ authors say generative AI hype has ‘spiralled out of control’ | VentureBeat, Sharon Goldmann writes about an interview she conducted with the authors of the ‘AI Snake Oil’ theory. Back in 2019, Princeton University’s Arvind Narayanan, a professor of computer science and expert on algorithmic fairness, AI and privacy, shared a set of slides on Twitter called ‘AI Snake Oil’. Snake oil is known as a product with no medical value. The presentation claimed that much of what is being sold as ‘AI’ is ‘snake oil’. In his view AI as applied today does not and cannot work. Narayanan who was recently named director of Princeton’s Center for Information Technology Policy, together with his Ph.D. student Sayash Kapoor, decided to explore what makes AI click, what makes certain problems resistant to AI and how to tell the difference. “When we started the research it was clear to us that generative AI – within a very short time – had made huge progress with millions of individuals exploring its usability. But we were caught off guard by the fact that it had become a consumer technology,“ says Narayan. The focus on AI has shifted from debunking predictive AI  to the application of generative AI, no matter what models you use and no matter how much data you have available. We are convinced that generative AI needs guardrails and more responsible handling by technologists engaged in government regulations. Better funding of our enforcement agencies would certainly help. People often think about AI-policy as a subject where one must start from scratch. However, that is not the case at all. Something like 80% of what needs to happen is just enforcing laws that we already have and to apply the regulations avoiding loopholes. In the last few months, there has been an increasing rift between the AI-ethics and the AI-safety communities. There is plenty of discussion that this is an academic rift which can be resolved because these communities are basically aiming for the same purpose. Many harmful things such as deep fakes are distributed over the internet. Hence, individuals need to be mindful of the potential misuse and they must use their collective power to decide how this technology should be used, whether at their workplace or their personal life.


According to the consulting company Gartner, generative AI and  GPT-4 have reached the peak of the so-called hype-cycle of expectations. The five phases of the Gartner Hype Cycle are defined as: Technology Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment and Plateau of Productivity. According to Gartner the main reason AI has hit this peak of Inflated Expectations, is the sheer number of products claiming to have generative AI baked into them. There is plenty of confusion in the market as vendors claim enormous productivity gains which are nowhere close to reality. Consequently, customers are still trying to reconcile the hype versus reality. Is this a technology that will provide long-term and sustainable value for enterprises and our society at large? Within a few years this question will be answered.


  • Hello Peter,
    great essay and synthesis;
    Natural language is not very precise (authors initial attempt, transfer to, receiver), same concern for derived methods such as AI.
    In addition the produced content redundancies are massive, ever wilder and overloaded. Pushed via many high tech tools, features available to anyone (Twitch TV streaming, consumer photo-video drones, geo VR navigation, social media posts size increases, DIY guides for functional medicine …. car repair).
    In particular when using BART (which uses also recent Internet information, bound to search) had recognised a strong overlap in the answers of familiar information. This minimal familiarities results in TRUST and engagement, dependencies. The overlap of familiar information has strong relation to Google high page rank (same, similar as in all AI models) which seems logical. The ‘page rank on Google, Tiktok,…’ has created/defined/disrupted many economies but does also result in the Manipulated Global Consciousness (now further inherited in many AI models).
    This week visited the International Visa Photo Journalism exposition in Perpignan also attending Conference session by Tyler Hicks / The New York Times.
    He outlined the massive change in past years and much reduced budgets for most but few news companies, many disappeared or republish content posted by others. The high misuse and propaganda during wars of internet information (his pictures in the Expo are from Ukraine war). Tyler Hicks being on personal/professional missions and TRUST being key to reach critical, dangerous places. He used his key term ‘TRUST’ several times.
    All exposed photographers stills outline world conflicts, challenges. Some stills difficult to view considering the captured despair but very essential as TRUSTED content. Encountered some emotional fatigue, depressive mood after viewing many stills. This was lightened-up when walking in the old town at noon with many outside restaurants fully populated, people enjoying the rest and food varieties (Going from the perceived World’s dark side to the bright side in a blink: all Visual). No or limited natural language was involved in this rich Expo experience.
    Yes, assume the AI hypes will practically adjust. Actually in Catalan schools the mobile phones are locket out some days per week to re-establish direct kids contacts between them via their many rich communication means, one of the natural language. The pitfalls, overkill with natural language are history, also well visible in the academic civil codes of countries.
    Your essays in natural language are much TRUSTed and excellent.
    best greetings

    • Dear Hanes,
      many thanks for your excellent comments about the issue of TRUST. It motivates me to use language as a means to communicate which cannot be replicated by LLMs or generative AI. Hence I am working on my next Essay which deals with the potential revival of analog computing.
      Herzliche Grüsse aus dem herbstlichen Meggen
      Peter Rudin

Leave a Reply

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