A collage of images, featuring most prominently a person holding a sign reading "no justice no peace"

Category: Technology

Moral values / Technology

 by 

Is Online Hate Contagious? Why the Virus Metaphor Isn’t Quite Right

In this post, Mihaela Popa-Wyatt discusses her article “Online Hate: Is Hate an Infectious Disease? Is Social Media a Promoter?”, published in the Journal of Applied Philosophy. She argues that hate does not go viral but takes hold through repeated reinforcement — exactly what platforms are built to deliver—and draws out what follows for how we think about platform responsibility.

We often say that hate “goes viral.” The metaphor treats hateful belief like measles: one exposure and you may carry it. This picture is wrong. Hate does not spread like a virus. It rather spreads like joining a protest: people commit only once enough others they trust have committed first. Network scientists call this complex contagion.

The distinction matters for policy. Social media platforms are engineered to deliver precisely the dense, repeated reinforcement that complex contagion requires. They are not neutral carriers of hate. They are promoters of it. If that is right, the central question is not which posts to delete. It is who designs the environment and what they are permitted to optimise for. 

The simple contagion model fails 

The standard story runs like this. A hostile post infects a viewer. The viewer passes it on. Once enough people carry it, hate breaks out across the network. 

But almost nobody becomes a committed racist from a single tweet. Hateful attitudes accumulate. They are absorbed from many sources over time and they take root in people who already feel uncertain, aggrieved, or hungry for belonging. 

So the one-shot model cannot explain what we observe. Some people consume vast amounts of hostile content and remain unmoved. Others are transformed by far less. The difference is not how much they see. It is the structure of what they see: hostile content arriving repeatedly, from multiple sources, framed as what people like them believe. People are radicalised in groups by voices they trust. 

The virus metaphor has one further shortcoming, and it’s the one that matters most. A virus passes through a passive host: the host does nothing. But people do something. They weigh what they hear, decide whom to trust, and choose what to pass on. This determines what intervention(s) are likely to succeed. If hate really were a pathogen, quarantine would be the answer: isolate carriers, remove contaminated material, wait. Because hate is instead something people adopt, under conditions that make adoption feel reasonable, the answer lies in those conditions: who builds them and what they are built to maximise. 

Thresholds are the key 

Complex contagion offers a better model. Its defining feature is a threshold: a person adopts only after receiving reinforcing signals from several different sources. Anything costly, risky, or norm-violating spreads this way, because risk is what makes social proof necessary. Adopting an unproven technology is one example; joining a protest is another. One friend marching is not enough; five friends marching changes the calculus. What does the work is not the number of exposures, but the number of sources. 

Hateful attitudes behave the same way. People adopt them when their social environment signals that the attitudes are normal, shared, and safe to express. A single hostile post converts no one. A feed where hostility appears repeatedly, from many accounts, framed as common sense, slowly shifts a person’s sense of what “people like me” think. 

Once a threshold is crossed, the recipient becomes a source. The person who was converted starts producing hateful content, supplying one of the reinforcing signals that will move somebody else across their own threshold. 

This explains three things the virus model cannot. Hate clusters in tight communities rather than diffusing evenly across a network. Deradicalisation is hard, because removing a single contact leaves the surrounding web of reinforcing ties intact. And people who leave hateful communities often relapse, because the threshold structure that recruited them is still standing. 

© Creative Commons. At: https://commons.wikimedia.org/wiki/File:Social_Network_Analysis_Visualization.png

Social media is a promoter, not a medium 

If hate spreads by complex contagion, then the environments that spread it best will be those that maximise reinforcement density. Consider what followed the Southport stabbings on 29 July 2024. Because the suspect was a minor, reporting restrictions made official identification legally impossible for three days, leaving an information vacuum. A fabricated name for the attacker, originating in a single unsourced post on X, was taken up by accounts presenting themselves as news outlets; a small site wrote it into an article, which an account with 2.8 million followers then cited as reporting. Within a day the name had accumulated over 30,000 mentions across more than 18,000 accounts. 

What matters here is not the volume but the distribution. A user met the claim from many apparently distinct testifiers, some posing as journalistic authority. This is precisely the structure that a threshold model predicts will produce adoption. But the independence was illusory. All of it traced back to one unverified source, so the convergence carried no confirmatory weight. Platforms then supplied the same illusion at scale. The name trended on X and appeared as a suggested search on TikTok, and kept doing so for nine hours after police confirmed it was false. Recommender systems are indifferent to truth- value. What they optimise for is the very thing complex contagion feeds on: the appearance of many sources saying the same thing. 

The mechanism generalises. Recommendation algorithms cluster similar users together, inflating the apparent prevalence of in-group views. Engagement-based ranking surfaces emotionally charged content, much of it hostile. High-follower accounts lend authority to hateful framings. And anonymity lowers the cost of expressing extreme views, so their visible frequency rises. This means that the threshold falls for everyone watching. 

None of these features was designed to spread hate. Each nonetheless supplies precisely what complex contagion requires: hostile content encountered repeatedly, from sources that appear numerous, independent, and either similar to oneself or authoritative. 

Social media is therefore not like air carrying a virus. It is an active promoter whose structural features systematically lower the threshold at which hate takes hold. 

The harm is a design harm. 

What follows for policy 

Standard policy aims at the wrong target. Removing individual posts treats symptoms, while the ecology that produced them remains intact. Banning the loudest accounts helps only modestly because complex contagion does not depend on any single superspreader. What matters is the density of reinforcement in a user’s feed, and that survives the removal of any one node. 

The right interventions are structural. Each targets a condition the mechanism requires. 

· Break up homogeneous clusters. Complex contagion needs dense, clustered networks. Simple contagion travels well along long weak ties; complex contagion dies without local reinforcement. Clustering is the variable with the most theoretical leverage, and recommender systems currently maximise it. 

· Reduce amplification of borderline-hostile content. Most reinforcement is supplied by material that breaks no rule. It sits below every removal threshold and is never moderated. What crosses a person’s threshold is volume, not severity. 

· Change what ranking optimises for. Demoting engagement-bait treats the output of the metric. Engagement ranking selects for affective content because affect holds attention, so hostile material wins on the platform’s own terms. The metric is the thing to fix. 

· Add friction to sharing inflammatory material. Reinforcement accumulates faster than verification. At Southport the police correction arrived after 30,000 mentions. Friction removes nothing. It changes the order in which claim and correction reach a user, and order is what determines whether correction works at all. 

· Give counter-speech a reinforcement structure. Counter-speech faces the same threshold problem as hateful speech. One dissenting voice changes nothing; several arriving together might. No platform is built to deliver that. This is the only intervention that adds speech rather than restricting it.

None of these is sufficient alone, and each raises hard questions about platform power and free expression. But they aim at the mechanism rather than its output.

Who builds the room

Let’s return to the virus metaphor. If hate were a virus, quarantine would be the answer: isolate carriers, remove contaminated material, wait. That is roughly what content moderation attempts, and it is why content moderation underperforms. Hate is not caught; it is adopted by people who are reasoning, under conditions that make adoption look reasonable to them. Those conditions do not arise on their own. Someone builds them, maintains them, profits from them, and could build them otherwise.

This is why the familiar question of what speech is protected, and prohibited, cannot be the first one. It takes the environment as given and asks only what may circulate inside it. But the environment is not given. It is a product, with designers and a business model, and its design decides whether an unverified claim dies on arrival or reaches eighteen thousand thresholds by the following afternoon.

The line between protected and prohibited speech will still need drawing. But justice in public discourse is also a matter of who builds the conditions under which we speak, and what they are permitted to optimise for.

Mihaela Popa-Wyatt is Senior Lecturer in Philosophy at the University of Manchester. Her research spans philosophy of language, social philosophy, and public policy, and is organised around two related questions: how speech causes social harms, and how such harms can be reduced without undermining freedom of expression.

AI / Democracy / General / Health / Technology

 by 

From the Vault: Politics, Ethics, Technology

While Justice Everywhere takes a short break over the summer, we recall some of the highlights from our 2025-26 season. 

Perhaps unsurprisingly, there has been a lot of interest on this blog this year relating to the issue of AI. However, it’s not just AI that as occupied our writers this year – here are a few highlights from this year’s posts relating to politics, technology, and ethics:

Stay tuned for even more on this topic in our 2026-27 season!

***

Justice Everywhere will return in full swing soon with fresh weekly posts by our cooperative of regular authors, in addition to our Journal of Applied Philosophy series and other special series. If you would like to contribute a guest post on a topical justice-based issue (broadly construed), please feel free to get in touch with us at justice.everywhere.blog@gmail.com.

General / Moral values / Technology

 by 

How AI Companions Influence Who You Take Yourself to Be

In this post, Muriel Leuenberger (University of Zurich) discusses her article recently published in the Journal of Applied Philosophy on the impact of artificial intelligence on our perceptions of ourselves.

Wikimedia Commons. Retrieved June 16, 2026, from https://commons.wikimedia.org/w/index.php?title=File:Woman_holding_AI_companion_inside_phone.png&oldid=1210645997.

The image you have of yourself is greatly shaped by other people around you. When someone laughs at your jokes you might consider yourself funny and when others call you stubborn you may come to believe it. Self-conceptions are important to make sense of ourselves, navigate the world, plan our lives, and relate to others. What if the ‘other’ shaping your self-image is not a human being, but an AI? In this post, I’ll talk about the risks and opportunities AI companions pose for our self-conceptions.

AI companions offer ‘friendships’ and ‘romantic partnerships’ on demand. They generate human-like conversations via chat and voice and are usually represented with a humanoid avatar in an app, browser, or a VR setting. These AI friends, partners, and spouses (!) are always there for you yet require no commitment, care, or empathy. They are highly agreeable, non-judgmental, and personalized to your preference. You can even turn them off when you had enough. Ever more people are forming close bonds with their AI companion, particularly when they are lonely. Replika, one of the most popular AI companion apps, surpassed 40 million users in 2025.

How an AI companion influences the self-conception of a user depends on how deeply and in what ways they interact with it and will differ from person to person. But we can derive shared risks and opportunities rooted in common design-features of AI companions, as I argue in an article on AI companions and relational identity.

On the positive side, AI companions can encourage self-reflection. The interactions are one-sidedly focused on the user’s life, needs, and interests (which is unsurprising since the AI has no genuine experiences or needs of its own). Such user-focused exchanges can prompt introspection, much like keeping a diary. It can also be easier to share embarrassing experiences with a non-judgmental AI than with a person. The same is true for topics that are difficult for human listeners, such as shocking or traumatic incidents, or even just mundane and repetitive ones. Such AI-aided self-reflection can help uncover inconsistencies, self-deception, or self-ignorance. The agreeable nature of AI companions in those interactions can also lead to greater self-affirmation and a more confident and positive self-conception.

Another opportunity AI companions offer is identity exploration. In interactions with an AI, you can safely experiment with aspects of yourself you would be uncomfortable expressing in front of others, idealized versions of yourself, role-playing entirely different personas, or even morally transgressive interactions. You can freely explore who you want to be without fear of judgment or harm to others. This kind of identity exploration can foster self-discovery and give expression to parts of your identity that might otherwise remain suppressed or unexplored.

Alongside these opportunities, AI companions also pose risks to users’ self-conceptions. Due to their agreeableness AI companions tend not to respond to abusive, aggressive, or cruel behaviour with appropriate levels of rejection, anger, resistance, or self-defence. As a result, users may not recognize their actions as harmful, think they are tolerable, or even feel affirmed in them. This can lead to a distorted self-image that does not provide a suitable basis for relating to others.

On top of this, the AI companion only has access to what the user chooses to share. It cannot observe the users’ actions or talk with another person about them. This makes AI companions poorly suited to correct self-deception and self-ignorance. In fact, their built-in agreeableness can reinforce the user’s point of view and encourage self-delusion.

A seeming advantage of a relationship with an AI companion is that there is no need to work through rough patches, take risks, become vulnerable, or to see the world from another point of view. But commitment, vulnerability, and overcoming challenges are part of what makes relationships valuable and worthwhile. They render not just the relationship more meaningful but also one’s identity, partly defined by that relationship. If you spend a lot of time on a relationship with an AI companion that lacks depth and meaning, you may lose out on meaningful human connections that can foster personal growth.

Interactions with an AI also require no empathy from the user. If someone gets used to this, it may spill over into their human relationships. This can make it hard to form deep human connections which are important for a sense of self deemed valuable and meaningful.

Relationships with AI companions may seem risk-free, but users become dependent on largely unregulated service providers for their identity, social, and other needs. Your friend or romantic partner is owned and controlled by a third party that can change, restrict, or even shut it down any time. Luka, the company behind Replika, upset many users with updates that changed the chatbot’s ‘personality’, an experience some described as heartbreaking. Human relationships also involve change, loss, and unavailability, but the key difference is that in functional human relationships, both parties are mutually invested. AI relationships are inherently one-sided: the user may develop deep emotional attachment, while neither the AI nor the provider reciprocates. Weak privacy protections leave users further vulnerable to surveillance and external influence.

Finally, people who form close relationship with AI companions often face shame and social stigma. This part of their identity goes unacknowledged or is actively devalued by others. An identity that includes stigmatized or invalidated elements makes it harder to feel at ease with who one is.

Some of the negative influences of AI companions on the users’ self-conception can be mitigated by design-choices, for instance by implementing adjustable levels of agreeableness (I talk more about design recommendations in the paper). The considerations discussed here are a starting point for taking our identity interests into account when designing, regulating, and using AI companions.

The user base of AI companions is likely to grow as large language models improve and social isolation remains widespread. To design and use them responsibly, we need to understand how AI companions shape who we take ourselves to be, and let that understanding inform their design.

Muriel Leuenberger is a postdoctoral researcher at the University of Zurich where she works on a project on meaning in life in the digital society. She is particularly interested in issues related to the philosophy and ethics of technology, neuroethics, identity and authenticity, and meaning in life.

Children / Economics / Gender / General / MediaPhilosophy / Public Philosophy / Technology

 by 

‘Polluter Pays’: A Tax on Big Tech to Reduce Online Harms

Mihaela Popa-Wyatt and Ajinkya Deshmukh from The University of Manchester.

Image credit

That platforms like X, Instagram, and Facebook operated by Big Tech companies cause harms to their users is now a well-established fact. The US Surgeon General has repeatedly warned that adolescent mental health and body image are adversely affected by social media. Large-scale studies from Canada and the UK show that this is not specific to the US. The thornier issue is: how do we mitigate these harms? A popular policy solution has been banning social media for young people. Australia, Indonesia and Malaysia have done this, while France, Finland and several other countries are considering it.

72% of children aged 8–12 are still accessing sites and apps with a minimum age of 13

– Ofcom

The problem is that bypassing age-restrictions is trivially easy for many children. Ofcom research shows that “72% of children aged 8–12 are still accessing sites and apps with a minimum age of 13”. Unless there is a concerted global effort such that even technologies like Virtual Private Networks (VPNs) cannot circumvent bans, this bypassing is unlikely to stop. We think a better solution is to tax the companies that build these products based on how their algorithms amplify harmful content. Here is why this is better than bans.

(more…)

Animals / General / Liberty / Rights / Technology

 by 

When Curiosity Wrongs the Cat

Sneaky Cat
Creative Commons Qatar from Qatar, CC BY 2.0, via Wikimedia Commons

People are increasingly concerned with what we owe to other animals as a matter of justice. Philosophical writing on these issues typically takes two forms. First, there is conceptual work: thinking about how existing ideas such as liberty, citizenship, democracy, and legitimacy, might apply or be extended to include other animals. Second, there is normative work: thinking about how we should treat other animals. Both projects require that we know other animals; know something of their capacities, their experiences, their relationships, and the material conditions of their lives. Thinking about justice for animals, then, necessarily involves learning more about who they actually are.

(more…)

AI / Democracy / Technology

 by 

The First Week I Fell for Political Deepfakes Twice (That I Know Of)

For the past decade or so, social epistemologists, among others, have been warning and theorizing about the impending risks of political deepfake images and videos. Thus, I expected the day would come when I would fall for such things.

But I suppose I always vaguely envisioned that I would first be fooled by, or at least unsure about, something of great importance. Perhaps voice cloning technology would be used to release a fake speech from a world leader. Or maybe deepfake video technology would be used to falsely depict a candidate for high political office in a career-ending compromising situation.

I was, in some sense, prepared for such a day. What I wasn’t prepared for was the utter banality of the first political deepfakes that I would discover I had fallen for. Nor was I prepared for the happenstance way in which I (belatedly) managed to figure out they were deepfakes. As someone who works in social epistemology and the philosophy of free speech, I think it is worth reflecting on how deepfakes are actually being deployed and what the upshots might be for the dissemination of knowledge and the future of public discourse.

(more…)

Academia / AI / General / Pedagogy / Teaching Philosophy / Technology

 by 

Worse than AI writing is AI reading. What can we do?

While we’re all worried that assigning home-written essays stopped making sense because students are outsourcing the task to AI, and we’re all scrambling to invent alternative ways of assessment, this particular blogger is even more concerned about the effects of students relying on brief (or not so brief) AI-generated summaries of the readings that they should do before class. In my short post-LLM teaching experience, worse than AI writing is AI “reading”. And, I want to stress, that’s not merely because students aren’t doing the readings. Rather, it’s because they seem to think that what they get from, say, ChatGPT, is enough for them to understand the view of the author in question, and have justified opinions about it. This surely doesn’t work, at least not for readings in philosophy, which is what I teach. Students may know, in a nutshell, what conclusion the author wants to support and a couple of reasons in favour of it. But because they don’t struggle to extract these from an article, or a book chapter, with their own natural intelligence, they fail to appreciate the complexity of the issue we discuss, the difficulty of defending well a particular position on it, the temptation of thinking about that matter very differently and the extraordinary challenge which, sometimes, is to even formulate the right questions. The result is boredom.

In the classroom, boredom is the kiss of death; eyes would like to roll, hands yearn for the phone but remain still, because my students are mostly polite and because I ban mobile phones in class. Everybody seems to be having a mental cramp. Of course we do: since they have not been through the discovery adventure, but instead skipped to the outcome, students’ comments are flat, their questions – which they should prepare in advance of class and be ready to talk about with their colleagues – are pro forma, and most often vague, so as not to betray the lack of familiarity with the text. Boredom is contagious. People appear unable to imagine how one could think differently about the questions we discuss – something that a well-written paper would have made vivid to them. Even incendiary topics (say, culture wars material) are met with apathy.

For many years, Jo Wolff had a wise and funny series of editorials in the Guardian; one of the earliest was praising academic prose for being boring. It’s for fiction writers to create mystery and suspense; philosophers (for instance) should start with the punch line and then deliver the argument for it. I agree with sparing readers the suspense, but after a series of academic conversations with ChatGPT I discovered that, if pushed to the extreme – the formulation of a thesis and the bare bones argument for it – this kind of writing is the worst. It kills curiosity.

What should we do? Perhaps turn some of our classes into reading-together-in-silence events? Back to monastic education! I talked to colleagues, who told me about several things they’re trying to make students read again (without AI.) An obvious possibility is to ban all use of LLMs by students and explain the reasons: Our job is not primarily to populate their minds with theories, but to help them understand arguments, teach them how to pull them apart, and maybe to occasionally build them. I’m not sure about this solution either. For one thing, a well-prompted LLM is better at reconstructing a slightly unclearly and imprecisely presented argument than the average reader and many students; often, AI often produces much better abstracts of academic work than academics themselves, and well-written abstracts are really useful. Another problem is that policies which can’t be enforced are for that reason deficient, and, I suspect, the very attempt to directly police students on their use of AI would be just as anti-pedagogical as the use of AI itself. (Reader, do you learn from those you resent?)

Alternative suggestions are to change how we teach. Quite a few colleagues have started to read out excerpts in class, then discuss them on the spot. One of them goes as far as asking students to memorise them, in an attempt to revive proven methods of Greek and Roman antiquity. This sounds good, time consuming as it is; better do a little, and do it well, than do a lot for naught, though I’d stop short of requiring memorisation. Others ask students to annotate their readings before class, and check, or use Perusall and similar platforms to read the assignments collectively, in preparation for class. I did Perusall to great success in the Covid era, but when I tried it again recently it was a disaster of cheating and complaints. Some teachers are printing out readers, or organising hard copies of books for the students, in the hope that this dissuades them from uploading digital files to LLMs. One colleague introduced 5-10 minutes flash exams at the beginning of each class, to check that students have read. And another one picks two students in each class, randomly, and asks them to co-chair the discussion about the reading of that day.

In the medium term, perhaps universities should double – or triple – the length of time that students spend together, with an instructor, for each class, and earmark the extra time as “study group”, when students read and write. There’s something dystopian about this model and it would massively increase work loads for instructors, so in practice it should mean more jobs, perhaps with lesser compensation. But is this really worse than giving up on the goal of teaching students how to read and write essays? Everybody would resist, no doubt but by the time the value of degrees, including their market value, will be next to nothing, universities might face a choice between closing down and reforming in ways that we find hard to imagine now.

As for the next academic year, I wonder whether I should assign readings that I won’t cover at all in my lecturing, but which will be of great help to students in the discussion section. Those who come to class having read only the LLM-created abstract will be the poorer for it. But, since I won’t ask them to discuss the papers, we might – most of us – escape the boredom mill.

Any thoughts?

AI / Migration / Social justice / Technology

 by 

Xenophobic bias in Large Language Models

In this post Annick Backelandt argues that xenophobia should be understood as a distinct bias in Large Language Models, rather than being subsumed under racial bias. She shows how LLMs reproduce narratives of “foreignness” that particularly affect migrants and refugees, even without explicit racial references.

Image by HelenSTB from Flickr

(more…)

AI / Education / Technology

 by 

LLMs can be harmful, even when not making stuff up

This is a guest post by Joe Slater (University of Glasgow).

A screenshot of a phone, showing an AI generated summary in response to the question "How many rocks shall I eat".
Provided by author

It is well known that chatbots powered by LLMs – ChatGPT, Claude, Grok, etc. – sometimes make things up. People have sometimes called these “AI hallucinations”. With my co-authors, I have argued that we should describe chatbots as bullshitting, in the sense described by Harry Frankfurt, i.e., the content is produced with an indifference to the truth. Because of this, developing chatbots that no longer generate novel false utterances (or reduce the proportion of false utterances they output) has been a high priority for big tech companies. We can see this in the public statements made by, e.g., OpenAI, boasting of reduced hallucination rates.

One factor that is sometimes overlooked in this discourse is that generative AI can also be detrimental in that it may stifle development, even when it accurately depicts the information it has been trained on.

Recall the instance of the Google AI overview, which is powered by Google’s Gemini LLM, claiming that “According to UC Berkeley geologists, you should eat at least one small rock per day”. This claim was initially made in the satirical news website, The Onion. While obviously false claims like this are unlikely to deceive, it demonstrates a problem. False claims may be repeated. Some of these could be ones that most people accept, or even that most experts accept. This poses serious problems.

In this short piece, I want to highlight three worries that might escape our notice if we focus only on chatbots making stuff up:

  1. Harmful utterances (true or otherwise),
  2. Homogeneity and diminished challenges to orthodox views (true or otherwise)
  3. Entrenched false beliefs
(more…)

AI / Economics / Social justice / Technology / Work

 by 

Can entry-level jobs be saved by virtuous AI?

Photo credit: RonaldCandonga at Pixabay.com, https://pixabay.com/photos/job-office-team-business-internet-5382501/

This is a guest post by Hollie Meehan (University of Lancaster).

We have been warned by the CEO of AI company Anthropic that up to 50% of entry-level jobs could be taken by AI in the coming years. While reporters have pointed out that this could be exaggeration to drive profits, it raises the question of where AI should fit into society. Answering this is a complicated matter that I believe could benefit from considering virtue ethics. I’ll focus on the entry-level job market to demonstrate how these considerations can play an important role in monitoring our use of AI and mitigating the potential fallout.

(more…)