The better algorithms of our nature

Engagement, bridging, and the design of digital platforms which don't pander to our weaknesses.

Via Tim Harford, I learnt of a new working paper on the emotions expressed when government policy is talked about. When the citizens and politicians talk about tax, immigration, or democracy, what emotions flavour that discourse? The authors developed an automated classification and looked at speeches by US politicians and posts by citizens on Twitter/X.

One emotion dominated all the others: anger

Emotions differed by topic. Discussions of gun control were more likely to feature fear, for example (shown in green in the figure below), but across all topics there was always one emotion that featured more than others, anger (shown in blue). On some topics (abortion, democracy) angry posts were more likely than posts which expressed no discernible emotion.

Figure 6 from “Emotions and Policy Views” (Algan et al, 2025)

Perhaps you are not surprised. Fear, anger and negative emotions generally are attention grabbing and provocative. No wonder, perhaps, that social media, like news, reflects this negativity bias. The paper reports that tweets expressing anger were more likely to be retweeted, meaning not only that anger is over-produced, but also over-spread.

An enduring human weakness for negativity, however, can’t explain patterns of year-on-year change.

The research team also tracked the change in anger over time, showing an increase between 2013 and 2025. This increase was from both the citizens’ responses to government policy (the “demand side” in econ-speak) and from politicians’ speeches and tweets (the “supply side” in econ-speak).

Two accounts for this could be a) non-angry people dropped off social media or b) the people who stayed got angrier. A ‘within-subjects’ analysis which looked at changes in expressed emotion by the same people over time suggests the drop-out account isn’t necessary to explain the observed data. On all policy issues the subsample they tracked across time got steadily angrier. On some, individual, issues, there was a discontinuous jump. Here’s a plot for anger in tweets on climate policy, which shows a dramatic increase post-2016:

Figure 10 from “Emotions and Policy Views” (Algan et al, 2025)

A lot happened in 2016. It would be easy to imagine that the entire reason for this change was the election of Donald Trump as president, and to stop looking for other causes, but something else also happened which I think we should at least consider, and that other thing says something important about how social media and human psychology interact.

In March 2016 Twitter completed their transition to the algorithmic feed, meaning that it was the default for all users to have their timeline populated by what Twitter thought they would want to see, rather than a chronological feed of posts from people they had decided to follow.

The algorithmic feed is now almost ubiquitous on social media. Exactly how these work for each platform is usually a guarded secret, but in general the major platforms use some form of engagement algorithm - meaning they try and predict what you will like, comment on, or even simply dwell on for longer than average. To do this they look at what you’ve liked, commented on, or dwelled on previously, as well as considering what people similar to you have engaged with.

Engagement algorithms have a nasty symbiosis with our human tendency to respond to threats. We’re already primed to pay attention to bad news, to pick up on other people’s emotions and respond in kind when people direct anger at us. Engagement algorithms give extra power to this negativity, since both hating something and loving it can equally look like strong engagement.

Crudely defined, engagement algorithms encourage expressing anger and the general polarisation of online discussion. Think of it this way. If there are posts along a spectrum of positions on an issue, say from left to right, posts all along the spectrum are likely to be liked by the different people who also align in their preferences from left to right.

All things being equal, you’d think this would mean the posts in the center had the most chance of attracting engagement, being able to recruit support from both sides. Sadly, we all know that modest takes are less fun to make, and extreme views are easier to articulate. Engagement algorithms which don’t distinguish a comment which says a post is completely right from a comment saying a post is completely wrong add to the advantage of the extreme ends of a spectrum. Viewed through the lens of an engagement metric, extreme contents get to count both its lovers and its haters towards their success.

Twitter users being cooked by the algorithmic fires of engagement-optimisation (actually a details from Fra Angelico’s Last Judgement showing the Souls of the Damned being tortured in hell).

The authors of the research paper are sanguine about the influence of the algorithm on the expression of anger on Twitter/X:

is the increase in anger mechanically due to the X algorithm? When X released the code of its algorithms in 2023, there was no direct mention of emotions or anger entering the ranking algorithm. However, the algorithm is designed to maximize retention and engagement so they may promote negative content not mechanically but because it generates more attention and, hence, revenues. This would not be contradictory to our story but instead show one mechanism through which anger can be perpetuated – because it causes more engagement.

So although they admit that the algorithm might be responsible for anger being promoted, they don’t mention the change of algorithm in 2016 as a possible source of the across-time changes. The role of the algorithm also undermines a simple interpretation of the greater retweet frequency of angry tweets. Are they retweeted because people are more likely to want to share anger, or because the algorithm puts those tweets in front of people and people can only share what they see? Both effects work in the same direction - people are either more likely to share angry posts, or the algorithm is predicting they are more likely to share angry tweets, or a mix of both. My guess is that the algorithm compounds an existing bias, adding to the polarising effect of the engagement algorithm counting violent disagreement the same as enthusiastic support.

§

So far I’ve just walked us through mechanisms which explain what we all know from experience — social media is often toxic, divisive and emotionally overwrought. What are the alternatives?

One is to get rid of non-algorithmic feed completely. This is the case with dissident social-media space Mastodon, part of a wider effort to produce federated, decentralised, non-commercial, social media (“the fediverse”). Often derided by those acclimatised to the hyperactive world of mainstream social media, there’s no algorithm on Mastodon. You see posts from people you follow, and only people you follow, in the order that they are made. The result is that Mastodon is harder work, slower and quieter. If I get 10 likes on Mastodon I consider that I’ve gone superviral.

The trick is understanding that Mastodon is not trying to replicate what Twitter used to be, or what X and other platforms are currently. This is a good post on that, but in short I’d describe the difference between mainstream social media and Mastodon as like the difference between trying to make friends in a banging nightclub and trying to make friends at a poorly attended fete in the village hall. There’s definitely more excitement and more people in the nightclub, but that’s not an unambiguous advantage to having a good time.

I used to posted regularly on Twitter, beginning in 2009, clocking up tens of thousands of posts and many thousands of followers by the time I stopped using using it in 2022. Since leaving I’ve never regretted it. I now devote my attention to Mastodon, like some retired veteran:

§

There is, though, an alternative to eschewing algorithms altogether.

Engagement algorithms, with their anger-encouraging and polarising effects, are just one option. The platforms we use are the result of specific design choices - every aspect built and rebuilt over time to influence behaviour.

Different choices could be made, different designs could bring out our better side rather than our worse.

For this, I’m excited by another class of algorithms called bridging algorithms. These identify and promote content that gets support across the spectrum of users. Bridging algorithms are used by the computational democracy platform pol.is to find consensus in policy discussions positions, and by the Community Notes content moderation system to identify when to put labels on controversial social media posts.

To understand the contrast with engagement algorithms, you need to know that bridging algorithms typically work with less information than engagement algorithms — they just look at likes or upvotes on content by users — but they do more than just count and add up user engagement. Instead, they use a statistical model which you can think of as trying to predict which users will like which content. By doing this, the algorithm can filter out engagement which is predicted by polarisation. So, for example, when left wing posts get liked by left-wing users or when right-wing posts get liked by right-wing users. The algorithm partials this out, identifying posts which get surprising endorsement; left-wing posts liked by right wing users, right-wing posts liked by left-wing users, etc.

The consequence of this method is that no camp can coordinate to promote their favoured view. It doesn’t matter how many likes a post gets, so there is no point left-wing or right-wing users coordinating to support a particular post. The algorithm will discount their votes, to the extent they are predictable based on their political leaning. This doesn’t mean that bridging algorithms can’t be manipulated, but it does make it far harder.

The overall intention is to select content which is endorsed regardless of partisan leanings. Not only are bridging algorithms neat, but they have value in providing a contrast to engagement algorithms.

We make the world, and there is an exercise in imagination to know that it could always be remade. Online too, we should plan and design to bring out our better selves. This isn’t to deny our weaknesses, including our susceptibility to negative emotions, but we can make choices about how the world will be fed back to us by the platforms and tools we use. If we’re going to rely on algorithms, we should use those which work with our better instincts instead of our worst.

---

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

See below for further reading other things I’ve been thinking about.

---

References

That working paper

• Algan, Y., Davoine, E., Renault, T., & Stantcheva, S. (2025). Emotions and policy views. Harvard University Working Paper. https://socialeconomicslab.org/wp-content/uploads/2025/07/emotions.pdf

Our paper on Community Notes:

• Arjmandi-Lari, Z., Mantzarlis, A., & Stafford, T. (2025). Threats to the sustainability of Community Notes on X. arXiv preprint arXiv:2510.00650.

Previously, on Reasonable People:

• The algorithmic heart of Community Notes

• Unhinged, unfortunate, unproven, inadequate, misdirected

• The Making of Community Notes

Other stuff:

the vocabulary of domestic recrimination

From the Origin Story podcast, this quote, which is from James Glover’s review of Carl Jung’s book which popularised the introvert/extrovert distinction :

“Jung’s type classifications contain an element of judgment which has caught the popular fancy and enriched the vocabulary of domestic recrimination.”

Origin Story: Introvert / Extrovert – In Two Minds

Indicator: Inside a pro-Conservative influence operation on Community Notes

As mentioned, Community Notes is the system on X (formally Twitter) which allows users to append notes providing context to posts. For example context like “The claims in this post are not true”. Alexios from Indicator reports a novel investigation which suggests a small group of users coordinated to prevent Notes appearing on posts from Conservative Party accounts.

The system is meant to make it hard for such coordinated manipulation to have an effect,but it is theoretically possible. My first observation is: the findings are compatible with this being an unofficial action by activists, a small group who took it upon themselves to try this, rather than a professional (and paid for) operation. Alexios details five accounts, two of which were only active during the 2024 UK election, so this might even have been 1 person, but probably no more than a handful. (of course with the advent of widely available agentic AI systems, I wonder what is stopping more professional actors infiltrating the Community Notes system with thousands of accounts).

My second observation is that this manipulation attempt pays a kind of perverse homage to both the importance of social media, and the threat of being seen not to tell the truth there. I am sort of amazed anyone is still on X to notice, but at least the Conservative Party thinks it matters what they post on there, and it matters that their posts don’t receive critical Notes.

Indicator: Inside a pro-Conservative influence operation on Community Notes

Paper: LLMpedia: A Transparent Framework to Materialize an LLM’s Encyclopedic Knowledge at Scale

Previously the factuality of LLMs has been assessed by querying them. This project generates an entire encyclopedia from model parameters, allowing an alternative — and more pessimistic — assessment of their accuracy

From the abstract:

For gpt-5-mini, the verifiable true rate on Wikipedia-covered subjects is only 74.7% -- more than 15 percentage points below the benchmark-based picture, consistent with the availability bias of fixed-question evaluation. Beyond Wikipedia, frontier subjects verifiable only through curated web evidence fall further to 63.2% true rate.

Paper: Saeed, M., & Razniewski, S. (2026). LLMpedia: A Transparent Framework to Materialize an LLM's Encyclopedic Knowledge at Scale. arXiv preprint arXiv:2603.24080. https://arxiv.org/abs/2603.24080

Site: https://llmpedia.net/

… And finally

Hybrid images encode two different pictures in different spatial frequencies. Non-technically, this means one picture is in sharp lines, another in blurred. The magic is that this makes a static image look different when viewed close up than when viewed from far away. If you don’t know what I mean, walk away from your screen looking at this image and see how it changes.

The same effect occurs when you view the image as big and small, which I realised this week is ideal for social media profile icons. Normally they are small, unless someone clicks on your picture to see the larger version. So I made a hybrid image to acts a memento mori surprise for anyone who clicks on my profile:

Small:

Same image big:

Again, you can prove to yourself it is the same image by walking further away from it.

The code to make these is here, if you want to play around https://github.com/tomstafford/hybridimages

Comments? Feedback? Hybrid images you’ve made? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.