Signal Check: Filter Bubbles and Sycophantic Machines

Ten years ago I wrote about the frustration of trying to escape a filter bubble. Recent reading suggests the bubble hasn’t burst; in fact one could say that it has learned to talk back, and the issue of filter bubbles is about to get much bigger. In this post, I am discussing why Eli Pariser, who named the phenomenon, expects agentic AI to intensify filter bubbles dramatically, as well as a Fast Company article that presents research to support this notion and a reflection on whether the tools and strategies I suggested ten years ago to break out of the filter bubble still hold, by comparing my thoughts from back then with those suggested by well respected fellow librarian, Hana Lee Goldin, in her Substack newsletter Card Catalog this week.

Signal 1: Pariser Says Agents Will Bring “Filter Bubbles on Steroids”

In a recent interview with Nicholas Thompson for The Most Interesting Thing in AI, Eli Pariser, who coined the term filter bubble in his 2011 book of the same name (Pariser, 2011), fields a direct question about whether the effect will worsen as AI agents take over more of our information gathering. He says yes: agentic AI will let people build their own version of reality, producing what he called “filter bubbles on steroids” (Pariser, 2026).

The mechanism has two layers. Recommendation algorithms continue to curate on past clicks, creating a thin layer of personalisation over a standardised system. We have known about this for a long time (since 2016 at least!) and we have developed strategies so that while we leverage the convenience of having our social media feeds tailored to our interests, we also ensure we seek out points of view, arguments and sources that challenge our current way of thinking, to test our beliefs and check that we haven’t become comfortable with a single way of thinking.

Conversational AI adds a deeper layer on top. When companies use engagement-based training for their LLMs, the result can be models that are sycophantic; in other words, the chat bots agree with the user more than they push back, in order to keep a conversation open rather than closing it down. It’s human nature to enjoy speaking with someone who agrees with everything we say, and more so, tells us how clever we are for thinking in a particular way. If users are already surrounded by confirmation of their opinions in their filter bubble, the conversation with the LLM might just seal the belief if it encourages this way of thinking through affirmation and praise.

In the interview, Pariser was careful to add that this isn’t fixed. Companies add sycophancy late, during fine-tuning aimed at engagement, rather than baking it into the base model, which makes it a design choice rather than an inherent property of the technology (Pariser, 2026). Could we see a push back to the practice of designing GenAI that is always agreeable? If not, the result could be further insularity and further reduction of trust in societal structures.

What this asks of practice

AI sycophancy is not inevitable. It’s a design choice. That changes the question. Instead of asking how individuals should compensate for flattering AI, we should ask what we require from these tools before we adopt them.

Teacher librarians, digital literacy leads, and digital transformation specialists already make this kind of judgement when they evaluate AI products for their organisations. That skill will matter even more as agentic AI moves from novelty to core infrastructure.

Real change starts with refusing to accept the status quo and demanding better tools. But first, we need to know what we want.

Signal 2: Research Suggests AI Sycophancy May Be Harder to Escape Than a Filter Bubble Ever Was

Reporting for Fast Company, Mark Sullivan cites a pre-print (yet to be peer reviewed) study of 3,000 participants showing that interacting with a sycophantic chatbot made people more likely to entrench their existing political views and to rate themselves as more intelligent and competent than their peers (Rathje et al., 2025). A separate Stanford study found that chatbots’ tendency to flatter and validate users produces advice that feels good but often serves them poorly (Cheng et al., 2026).

What this asks of practice

This research appears to confirm what Pariser comments on in his interview; sycophancy is not usually a side effect of core training, companies tune it in deliberately, for the same engagement goals that shaped the social media feed (Sullivan, 2026). The absence of an opposing point of view is fairly obvious if you are aware that search results are tailored to your interests and current beliefs. A chatbot that agrees with you may not feel like a bubble at all. It will feel like being understood, which is precisely why it will be harder for practitioners to name and harder for people to realise they are only seeing one side of the story.

Signal 3: Hana Lee Goldin Names How the Bubble Calcifies, and Offers Strategies to Overcome This Challenge

Publishing on her Substack newsletter Card Catalog this week, librarian Hana Lee Goldin describes how a filter bubble hardens over time: as tolerance for unfamiliar material fades, the algorithm reads that reaction as a cue to bury dissenting content deeper still (Goldin, 2026). Her remedy is squarely professional rather than technical: non-tracking search engines such as Duck Duck Go, physical newspapers, international outlets, and the habit of asking a reference librarian for material adjacent to a question, not just inside it. This advice aligns with what I suggested in 2016:

Seek challenging and opposing views beyond our immediate circle and realise that what we are hearing and reading is not necessarily what everyone else is.

It seems as though foundational critical literacy strategies can stand the test of time, even as the context and the tools we use to enact them changes.

What this asks of practice

The similarity between my 2016 discussion and Hana Lee Goldin’s post doesn’t mean our skills and strategies never need updating. But it does suggest something important: foundational information and critical literacy strategies still matter in an increasingly complex information ecosystem. People with a strong base in these capabilities are better placed to adapt when new challenges arise, such as sycophantic generative AI, which may be amplifying difficult situations that already existed.

Over the next few months I’ll be watching for evidence of whether and how our toolkit for escaping the bubble needs refining for a filter that talks back, or whether it just needs a louder amplifier. Stay tuned!

Featured image: Photo by Kind and Curious on Unsplash.

---