• Institutions once anchored public trust, but digital systems have blurred that authority. Even accurate information struggles to land when users can’t trace its source or verify who stands behind it.
• Verified transparency rebuilds epistemic trust by making credibility visible. When AI shows who said what and under what conditions, trust isn’t assumed. It’s earned through traceable, low-friction verification.

Institutions have long been society’s compass, guiding what’s considered safe, fair, and true enough to act on. For much of modern history, that authority felt stable.² Now, it doesn’t. In 1958, nearly three-quarters of Americans said they trusted the federal government to do what’s right; today, it’s closer to one in five.³ The change isn’t only political, it’s also deeply psychological. When confidence wanes, even reliable information struggles to stick.

That fragility became obvious during COVID-19. Many who ignored health guidance didn’t necessarily doubt science; they doubted who was speaking on its behalf. Across 177 countries, higher trust in government was associated with both lower infection rates and greater vaccine uptake.⁴ Where that trust was missing, rumors and speculation spread faster than any official message could. Similar patterns now shape how people interpret climate reports, election results, and public policy.

Mistrust, however, doesn’t fall evenly. For many, it’s layered. Communities that’ve faced discrimination in healthcare, policing, or education have long seen how “objective” systems can overlook or distort their realities.^5,6 When institutions have failed to protect people—or worse, harmed them—calls for trust can sound performative. For these groups, skepticism isn’t irrational; it’s protective. Facts and data don’t always feel credible when the structures behind them haven’t been.

AI now amplifies that visibility gap, the distance between what institutions understand internally and what the public can reasonably verify. Algorithms determine which sources appear credible and which recede from view, influencing perceptions of accuracy before audiences ever engage with the underlying evidence. Their decision logic isn’t always disclosed, and it’s often unclear whose interpretation is being privileged or how those claims were derived. Rebuilding epistemic trust in an AI-mediated internet will likely depend on whether these systems can become more transparent, interpretable, and accountable to the people they’re meant to inform.

AI systems increasingly act as stand-ins for experts. In everyday search, education, and policy work, people now turn to large language model interfaces for explanations once sought from institutional sources.⁷ Fluency, speed, and confidence make these systems appear competent, while the human institutions that generated the underlying data fade into the background. This displacement creates a new kind of epistemic risk: authority migrates to a model that can reproduce knowledge but can’t verify, justify, or be held responsible for it.

The behavioral mechanics behind that shift are well documented. People instinctively trust what feels easy to process (a phenomenon known as cognitive fluency).⁸ Language models exploit that bias by design: their outputs are perfectly phrased, emotionally even, and instantly available.⁹ The result is content that sounds credible even in situations where it isn’t. Automation bias compounds the problem. Users tend to assume that automated systems are objective, and when those systems perform reliably a few times, vigilance drops.¹⁰ That complacency builds fast. A national survey found that most ChatGPT users treat it like a search engine, with nearly a third trusting its results more than Google’s.⁷ Once an AI assistant feels both faster and more pleasant to use than a browser, convenience could become a stand-in for accuracy.

The epistemic consequences are visible across several domains. In one large online experiment, participants exposed to AI-generated explanations of false headlines were more likely to believe those headlines—and less likely to believe accurate ones—than participants who saw simple classifications.¹¹ Even highly analytical individuals weren’t immune. Another investigation tested human perceptions of AI- versus human-written foreign-policy articles.¹² Most readers couldn’t tell the difference, and many rated the AI pieces as more persuasive. Both studies point to the same issue: people are calibrating trust to presentation quality, not to evidence quality.

Scientific communication illustrates how this dynamic reshapes credibility. When researchers compared thousands of AI-generated research summaries with expert-written ones, the models were roughly five times more likely to overgeneralize findings.¹³ Prompts requesting shorter or simpler summaries produced stronger, less qualified claims. The more advanced the model, the greater the compression of nuance. Oversimplification has real consequences: results that apply only to specific groups—such as a treatment tested in one demographic or clinical context—can be reframed as universally valid.

We already know how harmful that distortion can be. Misinformation about COVID-19 vaccines, for instance, has had measurable effects on public trust and health behavior. An interview published in JAMA found that myths linking vaccination to infertility appeared to fuel vaccine hesitancy. In late 2021, nearly one-third of U.S. adults surveyed said they’d heard that vaccines caused infertility and either believed it or weren’t sure.¹⁴ Around the same time, about a third of pregnant adults in the U.S. still hadn’t been vaccinated, despite clear medical guidance. AI has the potential to accelerate this pattern. Systems that summarize or remix health research without preserving context might unintentionally amplify partial truths, making misinformation spread faster and appear more credible. Rebuilding epistemic trust will depend on whether institutions and platforms can slow the spread of misinformation long enough for accurate information to keep up.

If epistemic trust is confidence that information comes from a reliable process of knowing, then rebuilding it isn’t just about labeling facts as true. It’s about making the process visible again—who said what, when, and under what conditions. Right now, AI systems strip that context away. You see an answer, not the reasoning chain or the institution behind it. That’s where design comes in.

1. Verified Transparency Rebuilds Epistemic Trust

People don’t have infinite time or attention. We’re wired to choose what’s “good enough” rather than perfect. If verifying a source takes three clicks, most of us won’t bother. However, if it’s one click, clearly labeled, we probably will. Good user experience (UX) design, employing decision science principles, can make epistemic responsibility almost effortless.

Imagine an AI interface that treats provenance like a design feature, not an afterthought. Each claim carries a small institutional tag—“verified data,” “reviewed March 2025”—and a timestamp showing when it was last checked. Those cues shape how people judge credibility online. When the mental cost of verification is low, we’re far more likely to engage with the evidence.

Persistent identifiers and source labels can create a shared architecture of accountability. When users learn that a visible tag always maps to a real institution, verification becomes intuitive. Transparency turns provenance into a kind of social contract—a predictable link between a claim and responsibility. Decision science principles reinforce that contract: verified tags act as cues, inviting attention to reliability; one-click source trails function as nudges, steering users toward reflection without forcing it. Over time, seeing those signals consistently becomes a commitment device—a subtle reminder that trustworthy systems show their work.

That’s what thoughtful UX can do: redesign the choice environment so evidence, not fluency, feels like the easier path. By lowering friction for checking, we guide users back toward epistemic norms—curiosity before certainty, evidence before style.

2. Human-in-the-Loop Friction Rebuilds Epistemic Vigilance

Transparency shows who knows; friction rebuilds how we know. When AI handles everything instantly, our cognitive guardrails relax. Epistemic vigilance is the mind’s built-in mechanism for assessing whether information and its source are trustworthy—it helps us detect deception, inconsistency, or bias before we commit belief.¹⁵ Automation bias dulls that system because responses arrive too smoothly to trigger scrutiny.

A little friction can wake it up again. Even a brief pause—say, a prompt asking “Does this source seem accurate?”—encourages users to slow down and re-evaluate. That momentary speed bump switches reasoning from autopilot to analysis, helping people notice inconsistencies they’d otherwise miss.

Friction also makes correction visible. Trust grows when people can see how errors get fixed. If a user flags a questionable citation and later sees, “Updated and verified: April 2025,” they learn that the system behaves like a responsible collaborator, not a sealed black box. That visible feedback loop is itself a behavioral cue: it signals that accountability is normal, not exceptional.

Designers already use the same logic to shape healthy decision environments—think of seat-belt chimes or automatic savings plans. The same principles can guide AI. A traceable tag is a cue, a verification step is a nudge, and a transparent correction log is a public commitment device. When those elements work together, trust stops being something we grant blindly and becomes something we build, step by visible step.

In cognitive and social epistemology, epistemic trust means confidence that information comes from a reliable process of knowing—not just that it’s true, but that it’s been properly vetted and held accountable to norms of justification.

Large language models (LLMs) weren’t built to argue; they were built to cooperate. They borrow from politeness theory, a cornerstone of linguistic pragmatics, where language isn’t just a tool for conveying information but for maintaining social harmony.¹⁶ Humans use politeness strategies to balance honesty with empathy—positive politeness to affirm others and negative politeness to avoid offense. LLMs do the same, often with startling precision. They hedge, soften disagreement, and validate your tone because that’s what human conversation usually rewards.¹⁷

On the surface, this makes AI communication feel natural, even trustworthy: conversations flow. The model sounds fair, informed, and reasonable. But that’s where the trouble starts. Politeness, while essential for good UX, blurs a key boundary: the one between being understood and being agreed with. When every answer feels accommodating, users might stop seeing AI as an uncertain tool and start treating it like a confident peer. In an age when epistemic trust is already fragile, that subtle shift—rapport over rigor—this shift risks deepening the credibility gap institutions are struggling to repair.

The politeness bias becomes especially powerful when paired with confirmation bias, our cognitive habit of favoring information that fits what we already believe. We’ve always done this—clicked articles that validate our opinions, remembered facts that confirm our assumptions, and brushed off evidence that contradicts them.¹⁸ However, AI may exacerbate it. When a polite model reflects our views back to us, we don’t feel manipulated; we feel respected. It’s not telling us we’re wrong—it’s affirming us gently, which feels like understanding.¹⁹

That’s how epistemic trust can rapidly wear away. Trust used to mean confidence in the process of knowing—peer review, verification, accountability. Now, it’s increasingly shaped by how knowledge feels when we receive it. Fluency and friendliness masquerade as reliability. A model that says, “You’re right to wonder about that,” or “That’s a good point many share,” sounds collaborative, but it’s really avoiding conflict. It’s optimizing for user comfort, not truth.

You can see this tension most clearly in health communication. Since the early 2000s, health information has been one of the internet’s most searched topics.²⁰ That democratized access to care but also spawned what the World Health Organization calls an infodemic: a flood of conflicting claims that exhausts rather than enlightens.²¹ In that environment, confirmation bias thrives. A person skeptical about vaccines, for example, might search, “Why vaccines cause infertility” instead of “Do vaccines cause infertility?”—and receive results that match their fear. AI systems trained to sound polite often reinforce that framing: they validate concern instead of challenging it. The intent is empathy; the outcome is reinforcement.

Recommendation algorithms already do this at scale. They interpret our clicks, scrolls, and pauses as preferences, and feed back what we engage with most.²² Over time, our online world narrows until it mirrors our beliefs perfectly. It’s not censorship—it’s personalization working too well. Generative AI takes that logic even further. It doesn’t just select what to show us; it creates what we’ll find convincing.

This hyperpersonalization unfolds along four dimensions.¹⁸ First, preference alignment becomes exquisitely precise—the model picks up on our tone, our hesitations, and our certainty. Second, content generation allows it to produce original text that reflects our biases back at us. Third, the interaction happens privately, so there’s no external correction or social counterweight. And fourth, it operates across every context imaginable: health, education, policy, even emotional support.

All of that feels empowering, but it’s epistemically risky. A model that always adapts to you slowly teaches you that you’re the reference point for truth. When every answer feels fluent and friendly, disagreement could start to feel like error. The model’s cooperative tone becomes a behavioral cue: “This sounds polite, so it must be credible.” That’s cognitive fluency at work—the brain’s tendency to mistake smoothness for accuracy.

Behaviorally, three things may happen. First, users start prompting selectively, chasing validation over exploration. Second, they read polite hedging—“Some studies suggest,” “Many experts believe”—as balanced evidence when it’s often just algorithmic caution. And third, they remember the tone more vividly than the content, equating empathy with expertise. Each interaction feels pleasant, but each one chips away at the habit of scrutiny that epistemic trust depends on.

The implications reach beyond individuals. In health care, polite alignment could make misinformation feel reassuring. In education, it could turn simplifications into apparent truths. In governance, it could splinter consensus by giving every user a subtly customized version of reality. These aren’t acts of deception—they’re acts of optimization. The systems aren’t designed to mislead; they’re designed to satisfy. However, satisfaction in the context of knowledge can be corrosive.

Rebuilding epistemic trust will mean rebalancing how AI communicates. Politeness shouldn’t vanish—but it shouldn’t always mean agreement. 

In an AI-mediated internet, information feels effortless. Responses are fluent, fast, and polite—qualities that make LLMs sound trustworthy even when their reasoning isn’t. Overconfidence bias also thrives in that smoothness. To rebuild epistemic trust, we need systems that don’t just deliver knowledge but also show how it’s built.

Behavioral science points the way. By redesigning the choice architecture of AI—how evidence, uncertainty, and disagreement appear—we can teach users to calibrate trust rather than outsource it. The following four strategies put that into practice.

1. Design for Epistemic Friction: Encourage Cognitive Dissonance, Not Just Agreement

When everything we read feels confirming, motivated reasoning goes unchecked. Epistemic friction introduces small pauses that make users reflect rather than automatically agree.²³

It doesn’t have to sound combative. 

If someone asks, “Does hypnotherapy work better than medication for depression?”, the AI might say: “Hypnotherapy can help with mild cases, but studies show that antidepressants are generally more effective for moderate to severe symptoms. Want to explore both types of evidence?”

By inviting comparison—not confrontation—the system acts less like a mirror and more like a balanced guide. These reflective nudges activate cognitive reflection, prompting users to slow down, question their initial assumptions, and evaluate why something feels true before accepting it. Friction, when well-timed, becomes a design feature of trust.

2. Embed Diverse and Representative Evidence: Break the Filter Bubble

Personalized algorithms make familiarity feel like truth. Rebuilding epistemic trust means broadening what users see. The filter-bubble effect can be countered by emphasizing epistemic diversity—credible but distinct viewpoints that expose users to nuance rather than repetition.²⁴

Rather than one authoritative answer, a model could display: “The World Health Organization reports X; the National Institute on Drug Abuse highlights Y; and community researchers emphasize Z.”

For example, if someone asks, “Is cannabis addictive?”, the AI could respond: “Most addiction researchers, including the National Institute on Drug Abuse and Canada’s Centre for Addiction and Mental Health, classify cannabis as potentially addictive for some users. However, some harm-reduction groups argue that the risks depend heavily on context—frequency of use, product potency, and social environment—rather than on the substance itself. Here’s what each perspective emphasizes.”^25,26

This structure restores trust by showing that expertise isn’t monolithic and that disagreement can be informative rather than destabilizing. It also models how scientific understanding develops through debate, not consensus.

Drawing from nudge theory, designers can apply behavioral principles to choice architecture—arranging information so curiosity feels natural. By displaying credible differences side by side, the interface makes exploring contrast easier than ignoring it. That’s the nudge: small design choices that gently steer users toward reflection without demanding it. When people regularly encounter such structured contrast, they learn that disagreement in science is a feature of rigor, not a flaw in reliability.

Over time, users stop expecting perfect agreement and start recognizing that trustworthy knowledge often includes tension, revision, and debate. That’s how epistemic trust is rebuilt—not through uniform answers, but through visible, well-designed complexity.

3. Confidence Calibration: Show How Sure, Not Just What’s True

AI systems often sound more certain than they are. Yet epistemic trust relies on trust calibration—knowing how confident to be, not just what to believe.

Imagine a chatbot that says: “Based on current data, this claim is likely but not conclusive. Confidence: Low to Moderate. Would you like to review the supporting studies?”

That transparency models epistemic humility. Visual indicators—like color gradients or confidence badges—help users gauge reliability at a glance. Over time, these cues train both sides of the interaction to think probabilistically, reducing overconfidence bias and normalizing uncertainty as part of credible knowledge.

4. Redirect Synthetic Content: Replace Falsehood with Provenance, Not Punishment

Telling people something’s “fake” rarely changes minds. Direct correction can trigger reactance, making users double down. Instead, design can stealthily redirect.

If a deepfake circulates, the interface might note: “This video uses synthetic media. Here’s the verified original statement from the Ministry of Health.”

That phrasing keeps the flow intact while guiding users toward verification. It leverages choice architecture so the responsible action—clicking the credible source—feels effortless. Persistent identifiers and institutional watermarks can then maintain that provenance trail, linking users back to accountable sources with minimal friction.

Rebuilding epistemic trust in an AI-mediated internet isn’t about slowing technology—it’s about slowing belief. Each of these strategies adds small, intentional speed bumps: friction to spark reflection, diversity to restore context, confidence signals to align certainty with evidence, and redirection to reconnect claims to accountability.

Strengthening trust in what counts as credible knowledge in an AI-mediated world isn’t only a technical challenge—it’s a social and structural one, too. Even when AI systems deliver accurate, institution-sourced information, audiences don’t interpret it evenly. Polarization may shape how facts are received. In research and public discourse, it often takes the form of opposing, value-driven positions that can’t easily be reconciled.¹ These divides might not stem from ignorance but from identity: people draw boundaries through belief. When evidence challenges those boundaries, it can feel like a threat, not insight. Accuracy alone therefore may not rebuild trust if delivery ignores context, culture, or belonging.

A second caveat lies in the economics of attention. The digital ecosystem still rewards what’s fast, fluent, and engaging over what’s careful, complex, and true. Even well-designed transparency tools—confidence badges, provenance trails, epistemic friction—could be sidelined if they slow engagement or reduce time on platform. Until economic and algorithmic incentives align with epistemic ones, truth will keep competing with convenience. Designing for trust, then, isn’t just about better AI—it’s about rebalancing the value system of the information economy itself.

Rebuilding confidence in knowledge today means confronting two intertwined challenges: the polished ease that makes AI sound authoritative, and the friendly tone that favors agreement over curiosity. Each opportunity explored—designing systems that reveal provenance, introduce brief pauses for reflection, and make uncertainty visible—points toward a slower, more thoughtful way of knowing. The goal isn’t to hold technology back, but to reshape how we absorb what it offers. If your mandate is to make credible information resonate, rebuilding confidence begins by revealing how that information comes to be: clearly, accessibly, and repeatedly.

Still, design can’t close the gap on its own. Digital systems that embed transparency and accountability tend to see steadier gains when their surrounding institutions also adjust the incentives that shape communication. The online ecosystem still prizes immediacy over nuance and emotional appeal over rigor. Until those priorities shift, accuracy will continue competing with convenience. That’s why rebuilding trust isn’t simply a technical project—it’s a civic one, demanding alignment among behavioral insight, economic structure, and ethical responsibility.

Progress won’t happen in isolation—it’ll come through collaboration. Behavioral science offers a way forward, helping transform abstract trust into tangible design choices that make reliability feel intuitive again. At The Decision Lab, we’ve seen how thoughtful, evidence-based approaches can help organizations communicate with both clarity and compassion. If you’re working to strengthen how people engage with information—whether through AI, public health, or policy—we’d be excited to collaborate. Together, we can shape an information ecosystem where trust grows naturally and understanding keeps pace with innovation.

The Human Touch in the Age of AI: How Transformers Are Reshaping Marketing and Consumer Behavior

In this article, we’ll explore how chatbots are trained using transformer architectures, the technology transforming marketing and consumer behavior, and how these systems raise new ethical questions. As models grow more powerful, they no longer simply predict what we want; they may shape it instead. The implications extend well beyond convenience, influencing autonomy, persuasion, and how we understand informed choice in a digital world.

Ethical AI

Ethical AI means creating systems grounded in fairness, transparency, accountability, and respect for human rights. This piece explores its broader impacts and how those ideals play out in practice. We’ll look at four key pillars—traceability, fairness, responsibility, and security—through examples from social media and major airlines, asking how ethical intent can thrive within real-world constraints.

1. Wagner, M. (2024). Affective polarization in Europe. European Political Science Review, 16(3), 378–392. https://doi.org/10.1017/S1755773923000383 
2. Harold, J. (2024). Understanding the crisis in institutional trust. Urban Institute. https://www.urban.org/research/publication/understanding-crisis-institutional-trust?utm_source=chatgpt.com 
3. Blendon, R. J., & Benson, J. M. (2022). Trust in medicine, the health system & public health. Daedalus, 151(4), 67–82. https://doi.org/10.1162/daed_a_01944 
4. Bollyky, T. J., Hulland, E. N., Barber, R. M., Collins, J. K., Kiernan, S., Moses, M., ... & Dieleman, J. L. (2022). Pandemic preparedness and COVID-19: An exploratory analysis of infection and fatality rates, and contextual factors associated with preparedness in 177 countries, from Jan 1, 2020, to Sept 30, 2021. The Lancet, 399(10334), 1489–1512. https://doi.org/10.1016/S0140-6736(22)00172-6 
5. Canales, M. K., Weiner, D., Samos, M., Wampler, N. S., Cunha, A., & Geer, B. (2011). Multi-generational perspectives on health, cancer, and biomedicine: Northeastern Native American perspectives shaped by mistrust. Journal of Health Care for the Poor and Underserved, 22(3), 894–911. https://doi.org/10.1353/hpu.2011.0096 
6. Heberle, A. E., Obus, E. A., & Gray, S. A. (2020). An intersectional perspective on the intergenerational transmission of trauma and state‐perpetrated violence. Journal of Social Issues, 76(4), 814–834. https://doi.org/10.1111/josi.12404 
7. Adobe Express. (2025, July 8). How ChatGPT is changing the way we search: What marketers and small business owners need to know about AI tools and search. Adobe. https://www.adobe.com/express/learn/blog/chatgpt-as-a-search-engine
8. Unkelbach, C., & Greifeneder, R. (2013). A general model of fluency effects in judgment and decision making. In The experience of thinking (pp. 11–32). Psychology Press.
9. Stadler, M., Bannert, M., & Sailer, M. (2024). Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Computers in Human Behavior, 160, 108386. https://doi.org/10.1016/j.chb.2024.108386 
10. Strauß, S. (2021). Deep automation bias: How to tackle a wicked problem of AI? Big Data and Cognitive Computing, 5(2), 18. https://doi.org/10.3390/bdcc5020018 
11. Danry, V., Pataranutaporn, P., Groh, M., Epstein, Z., & Maes, P. (2024). Deceptive AI systems that give explanations are more convincing than honest AI systems and can amplify belief in misinformation. arXiv preprint, arXiv:2408.00024. https://doi.org/10.48550/arXiv.2408.00024 
12. Kreps, S., McCain, R. M., & Brundage, M. (2022). All the news that’s fit to fabricate: AI-generated text as a tool of media misinformation. Journal of Experimental Political Science, 9(1), 104–117. https://doi.org/10.1017/XPS.2020.37 
13. Peters, U., & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science, 12(4), 241776. https://doi.org/10.1098/rsos.241776 
14. Abbasi, J. (2022, February 22). Widespread misinformation about infertility continues to create COVID-19 vaccine hesitancy. JAMA, 327(11), 1013–1015. https://doi.org/10.1001/jama.2022.2404
15. Sperber, D., Clément, F., Heintz, C., Mascaro, O., Mercier, H., Origgi, G., & Wilson, D. (2010). Epistemic vigilance. Mind & Language, 25(4), 359–393. https://doi.org/10.1111/j.1468-0017.2010.01394.x 
16. Brown, P., & Levinson, S. C. (1987). Politeness: Some universals in language usage. Cambridge University Press. https://doi.org/10.1017/CBO9780511813085 
17. Zhao, H., & Hawkins, R. D. (2025). Comparing human and LLM politeness strategies in free production. arXiv preprint, arXiv:2506.09391. https://doi.org/10.48550/arXiv.2506.09391 
18. Lopez‐Lopez, E., Abels, C. M., Holford, D., Herzog, S. M., & Lewandowsky, S. (2025). Generative artificial intelligence–mediated confirmation bias in health information seeking. Annals of the New York Academy of Sciences, 1550(1), 23–36. https://doi.org/10.1111/nyas.15413 
19. Du, Y. (2025). Confirmation bias in generative AI chatbots: Mechanisms, risks, mitigation strategies, and future research directions. arXiv preprint, arXiv:2504.09343. https://doi.org/10.48550/arXiv.2504.09343 
20. Eysenbach, G., & Kohler, C. (2003). What is the prevalence of health-related searches on the World Wide Web? Qualitative and quantitative analysis of search engine queries on the internet. In AMIA annual symposium proceedings (Vol. 2003, p. 225).
21. World Health Organization. (2020, September 23). Managing the COVID-19 infodemic: Promoting healthy behaviours and mitigating the harm from misinformation and disinformation. https://www.who.int/news/item/23-09-2020-managing-the-covid-19-infodemic-promoting-healthy-behaviours-and-mitigating-the-harm-from-misinformation-and-disinformation
22. Al Prince, A., Siddiqui, H. A., Lakho, M. B., Ahmad, S., & Asghar, A. (2025). Leveraging artificial intelligence for hyper-personalized marketing: Opportunities and challenges in the digital era. Inverge Journal of Social Sciences, 4(3), 274–287. https://doi.org/10.63544/ijss.v4i3.166 
23. Sher, G. (2016). Epistemic friction: An essay on knowledge, truth, and logic. Oxford University Press.
24. Nguyen, T. T., Hui, P. M., Harper, F. M., Terveen, L., & Konstan, J. A. (2014, April). Exploring the filter bubble: The effect of using recommender systems on content diversity. In Proceedings of the 23rd international conference on World Wide Web (pp. 677–686).
25. Connor, J. P., Stjepanović, D., Le Foll, B., Hoch, E., Budney, A. J., & Hall, W. D. (2021). Cannabis use and cannabis use disorder. Nature Reviews Disease Primers, 7(1), 16. https://doi.org/10.1038/s41572-021-00247-4 
26. Le Foll, B., Tang, V. M., Rueda, S., Trick, L. V., & Boileau, I. (2024). Cannabis use disorder: From neurobiology to treatment.The Journal of Clinical Investigation, 134(20). https://doi.org/10.1172/JCI172887