Envisioning Information Access Systems: What Makes for Good Tools and a Healthy Web?

2.1 Definitions

The following is a list of a few of the most important concepts in the space. The relationships between them are also conceptually represented in Figure 1. It represents the view, widely accepted by scholars (e.g., [25, 43, 90]), that IB includes all types of interactions with information that a person has, some of which could be intentional (IS) and some not. Within those intentional behaviors, some may result in information being accessed—through various processes, including searching, scanning, and browsing.

Fig. 1.

View Figure

Information Behavior. IB refers to people interacting with information in a given context [91]. Note that this involves not only the information that people need or seek but also what they encounter accidentally and serendipitously.

Information Seeking. IS refers to the case where a person seeks information they have realized a need for [55]. Here, intentionality matters. IS refers to an intentional action by a person and does not entail successful IA: just because they are seeking information that does not mean that they find it or that the information even exists.

Information Retrieval. IR refers to a process through which the information sought by a user is found and typically ranked or organized in some way, if there are multiple possible relevant pieces of information or sources [73]. As a subset of IS, IR deals with cases where we believe the information being sought exists and focus on building processes and systems that ensure its proper retrieval and presentation. In other words, while IS is a user-driven and user-focused area, IR often tends to focus more on those processes and systems. Of course, there are gradations here. For instance, interactive IR (IIR) is where scholars focus on IR with much more involvement of users and retrieval through interactions, often providing more focus to the users [72].

Information Filtering. Information Filtering (IF) refers to a system being able to present relevant information to a user without the explicit or expressed need, such as a query or a question. Recommender systems fall under this category, as they typically use attributes of available items or the past behaviors of their users to suggest information, with the former being the case of content-based filtering and the latter being the case of collaborative filtering [70]. A long time ago, Belkin and Croft [12] examined the seeming duality of IR and IF—calling them two sides of the same coin.

Information Access. IA refers to a focused interaction between a person and information where relevant information is sought, found, and used—with or without a system [23]. This covers, but is not limited to, IR and IF, and is the focus of this article. We will examine this area further in the next subsection.

2.2 Types of IA

One can access information in two primary ways: directly or with algorithmic mediation. Visiting a website or scanning a downloaded document for relevant information are forms of direct access. Using a search engine that provides a ranked or organized set of results and browsing through recommendations on a media app are examples of IA that are algorithmically mediated.

Key elements to consider here include agency and transparency. In case of direct IA, the user has the most agency and transparency. Both of these decrease with the case of search engines (IR). The retrieved information that the user can access is presumably relevant to the submitted query, but it may not be very clear to the user why only these results were returned nor on what basis they are ranked. In case of recommendations (IF), this transparency is even further reduced due to the lack of a seed query or request stemming from the user’s need. We might conceptualize this as a tradeoff: accepting lower transparency in exchange for access to information sources we did not know to ask for. But we might then also ask if there are other avenues to transparency in such systems.

Before modern search engines and online IA systems, people relied on librarians and other subject or search experts to provide them with results and recommendations. While such experts may not always reveal why and how they found or suggested some information, there were and are typically well-accepted and well-understood models for how their processes work. Examples include the Dewey Decimal System [62], reference interview guides [29], and sense-making inquiry [28].

Note that in all of these cases, while the user loses some level of agency going from direct IA to mediated access, they still retain enough of it to be able to drive the process, question the outcomes, and find alternatives. All of these are put in jeopardy as we look at the newest and increasingly popular form of IA: generative IA. By definition, this falls under algorithmically mediated IA, but while typical algorithmic mediation happens for matching, ranking, and organizing existing information, generative IA involves synthesizing text based on information about word distributions and sometimes fine-tuning with supervised training data pertaining to human preferences. The result is text that might even be taken as new ‘information,’ despite the fact that the systems lack any world model or understanding of the text they manipulate [14].

To support a more thorough investigation of tradeoffs among desired properties of IA systems, we need to first explore that those desired properties are. We take up this topic next.

4.1 Assessment of IA Systems along the User-Focused Dimensions

In Section 3, we identified several characteristics that users want from an IA system, with six of them elaborated: relevance, novelty, speed, personalization/contextualization, interactivity, and transparency. Let us now examine how discriminative and generative IA systems do on these six dimensions. Note that both these categories fall under algorithmically mediated IA. Discriminative systems classify or rank existing content, whereas generative systems create new content based on the underlying LLM trained on large corpora. Both of these categories of IA systems cover IR (search) and IF (recommender systems) as shown in Figure 1. A quick comparison of discriminative and generative systems is presented in Table 1.

As far as relevance goes, both types of systems are able to produce impressive results, but the path to achieving that relevance varies. Search engines and other discriminative models achieve relevance through matching candidate retrieval results to the input queries and then ranking (although topical relevance is not the only factor in either of these processes). Generative models, however, achieve relevance either by running a discriminative process first and then using the generative model to synthesize a summary or just through their text synthesis process: with sufficiently large models, the training objective of plausibility will keep the text largely on topic and thus relevant. Similarly, novelty is produced in very different ways: for discriminative models, it is a question of what is available and how it is ranked. For generative models, it is dependent on pre-training and fine-tuning, and risks producing something that is both novel and false, through stochastic generation [13]. One of the interesting ways the notion of novelty is manifested with generative models is through the ‘temperature’ parameter. This parameter can allow the user to control how creative the model should get during content generation. For example, Bing provides three levels of conversation styles: Precise, Balanced, and Creative, going from more factual and retrieval-based output to more creative or novel content. Even though the user gets to control such temperature or style parameters, it is presented without thorough documentation. For example, a user may think that using the setting precise (temperature=0) would guarantee accurate answers, but of course this is not how LLMs work.

Both discriminative and generative systems provide opportunities for users to interact with the information surfaced, but in different ways: classical discriminative systems facilitate user interaction with the source documents. These documents themselves are typically static, but the user can examine them directly and also explore how they are situated (where they are hosted, what other sources of information they point to in turn, etc.). Generative systems provide a new kind of possibility for interaction, namely conversational chat. However, this comes at the cost of direct access to sources. In the scenario where answers are provided directly from an LLM trained on a large dataset, there either is no source for the information (it is a recombination of words or word parts from the training data that does not exist in its entirety or match the information in any source document) or it is not traceable.³ In the scenario where the LLM is used to summarize information from a duly linked set of source documents, the chat interface still seems likely to discourage exploration of those source documents, by foregrounding an appealing even if possibly incorrect summary.

Comparing discriminative and generative systems, we also see a tradeoff between personalization/contextualization and transparency. LLMs are masters of mimicry and can be prompted to use many different styles of language, for example. An LLM chatbot-based IA interface could have preset (‘hidden’) addenda to prompts⁴ along the lines of ‘Provide the answer using simple language’ that might be effective at producing output perceived as easier to understand for specific audiences (e.g., children, second language learners). However, this comes at the cost of the interface consisting solely of synthetic text. Furthermore, the more need there is for something like text simplification, the less well positioned the user would be to check the accuracy of the output they see. Similarly, such prompt addenda could possibly be used to help populations experiencing oppression (like LGBTQ+ users mentioned in Section 3) access output that contains less of the discriminatory framing that results from general searches might contain. In this case, a good point of comparison in studies of user satisfaction would be search engine techniques that help users find community forums and similar spaces where other people with similar experiences share knowledge—but note that shaping query responses to avoid discriminatory language, while potentially quite desirable, is a far cry from providing connections to community.

4.2 Potential and Projected Harms and Costs for LLM-Based IA Systems

In the current narrative around ‘generative AI’ and specifically LLM-based models, several issues have been raised [13, 77, 88] about their potential costs and harms. In the following, we examine a selection of these issues, starting with the most immediate and common issues and going toward longer-term harms and less talked about problems.

Ungrounded Answers. Despite all of the hype, from industry labs and elsewhere, that LLMs are not only a step on the path toward ‘artificial general intelligence’ [2], but in fact showing the first ‘sparks’ of it (phrasing from the title of the unscientific report released from Microsoft Research [22]), there is in fact no evidence or reason to believe that ‘intelligence’ or ‘reasoning’ will emerge from systems designed to extrude plausible sequences of word forms. With a large enough corpus for training, such a system can find sufficiently rich associations among words to generate appealing passages of text. But the only information an LLM can be said to have is information about the distribution of linguistic tokens [13, 14]. Just because a system is able to store and connect some representation of online texts does not make it knowledgeable or capable producing reliable answers. Information is not knowledge, and that is even more true when the information is only information about the distribution of word forms.

Bias in Answers. Friedman and Nissenbaum [38] define bias as systematic and unfair discrimination. We know that any IA system can represent and reproduce biases resulting from data, algorithms, and content presentation [8, 65]. And LLM-based generative systems could make things even worse: it is well established that LLMs absorb and amplify biases around gender, race, and other sensitive attributes [1, 16, 17, 18, 53, 78]. What is worse is that when these biases are reproduced in the synthetic text output by the machine, they are stripped of their original context, where they are more clearly situated as the ideas of people, and reified into something that seems to have come from an omniscient and impartial machine. This effect is not new—Noble [65] showed how Google’s presentation of search results had the same effect—but we believe there is reason to fear it will be even more pernicious when machines seem to be speaking our language.

Lack of Transparency. While most algorithmically mediated IA systems lack transparency, the issue becomes amplified when the system is LLM based and generating responses without a clear indication of how it sourced or created such responses. What data or information was used to train it? What was not used and why? Why was a given response generated? What signals and sources did it use? What confidence or guarantee can it provide? We should also consider the connection between transparency and accountability: if the response is incorrect, what recourse would be available to the user? What recourse would be available to non-users about whom incorrect information was displayed?

Lack of Agency for Users. As described earlier, one of the issues with algorithmically mediated IA is that the user often does not have enough agency. This is amplified when the system eliminates the interface with multiple options (e.g., set of search results on a page, or a set of recommendations on a widget) and provides a single response. It becomes difficult to impossible for the users to confirm or control these responses. They are not able to question or change the process through which their responses are generated, other than specifying, via prompt engineering, the type or length of the response they want. They are further not able to locate the sources in a broader information ecosystem and integrate that into their sense-making process [77].

Lack of Appropriate ‘Friction.’. Most systems are designed with the idea that they are supporting users who prefer to put in as little effort as possible when accessing information. This is certainly a desired characteristic of an IA system. However, we argue that it is not always advisable to minimize this effort, especially if it takes away the ability for a user to learn, for instance due to lack of transparency. We believe, and as others have supported [68], that certain amount of ‘friction’ in an IA process is a good thing to have. It allows the user to understand and question the process, as well as providing an ability to refine and even retract their original question or information need. LLM-based generative IA systems often go too far in reducing the active participation of the user as well as potentially beneficial friction as they aim to cut the process of discovering information to simply getting ‘the’ answer.

Labor Exploitation and Environmental Costs. Apart from the issues discussed previously arising from the use of LLM-driven chatbots, there are also serious issues with the way that they are presently produced. The performance of LLMs relies on extremely large datasets which are collected without consent, credit, or compensation [57].⁵ Current practice in improving the ‘safety’ of models like ChatGPT (read: reducing the chances that they output harmful content) involves a technique called reinforcement learning from human feedback [67], a data-intensive task which requires a human workforce to label frequently extremely troubling data. Investigative reporting by Billy Perrigo, Karen Hao, Josh Dzieza, and others has found that this work is largely outsourced to poorly paid workers in locales such as Kenya⁶—a finding in keeping with what is known about the human labor behind so-called ‘AI’ more generally [89]. Finally, there is the environmental impact of these compute-intensive systems (at both train and test time), which includes not only energy usage (and the associated carbon impact) [20, 76, 81] but also intensive water usage.⁷

4.3 Potential Beneficial Use Cases

There have been plenty of applications of generative IA systems in the recent months, including in education [10], healthcare [50], and commerce [94]. However, insufficient attention is given to making such applications both beneficial and safe. We argue that to be beneficial and safe, such a use case would have to be one where:

Even setting aside conditions (B) through (D), what kind of IA use case could satisfy (A)?

The most compelling use cases of ChatGPT that we have seen reported are for accessing information about how to express certain programs in different programming languages [80]. We note that this use case is actually best understood as a kind of machine translation, which is different from open-ended conversational replies. The answers (suggested computer code) are grounded in the questions (descriptions of what the computer code should do). Furthermore, the accuracy of the output can be relatively thoroughly checked, by compiling and running the code. We note, however, that issues remain, such as the possibility of security vulnerabilities in the generated code and lack of clarity around licensing/copyright of the training data [33]. Furthermore, as discussed further in Section 6.1, StackOverflow found that users with easy access to a system that provided answers that look good did not reliably test them before passing them along, causing headaches for the site.

Another possible use case⁸ involves tip-of-the-tongue phenomena: cases where the user is looking for the name of something like a movie or an obscure term, and can describe it. In this case, the results can be immediately verified with an ordinary search on the retrieved term. Condition (B) is certainly met as well. Condition (C) remains an issue: the chatbot might well output either something directly toxic or an answer that, when paired with the query, reflects damaging biases (akin to the examples discussed in Section 6.1). And condition (D) remains unresolved. Does a handy machine for turning to with tip-of-the-tongue queries merit the environmental and social impact of its construction?

4.4 Do LLMs Belong in IA Systems at All?

A critical perspective on LLMs in IA systems is open to the possibility that the tech simply does not fit the task. Accordingly, in this section we ask whether LLMs belong in IA systems at all.

In answering this question, the first step is to distinguish among ways that the LLMs might be used. Previous generation LLMs, such as BERT [30], were largely used as a source of ‘word embeddings’ or vector-space representations of words in terms of what other words they co-occur with. These representations are more informative than the word spellings themselves and thus proved extremely beneficial in many natural language processing classification tasks [61]. Classification tasks that serve as components of IA systems include named entity recognition, topic clustering, and disambiguation, among others. This use of LLMs does not seem to reduce user agency or transparency.

LLMs might also be used in query expansion, rephrasing the user’s query into multiple alternatives that can then in turn match more documents [9, 96]. This level of indirection comes at some cost to transparency (although systems could presumably make the expanded queries explorable). However, it still leaves the user with the ability to directly interact with the returned documents in their source context.

Finally, LLMs could be used to synthesize answers to user queries, either based on a specific set of documents returned (as done, e.g., in Google snippets [82]) or off of their entire training data (as with ChatGPT [66]). While the latter case can be called ungrounded generation, the former is aimed at being grounded on relevant content. Some of the recent efforts related to retrieval-augmented generation provide potential solutions for provenance and precision or accuracy but are still vulnerable to the shortcomings of text generation [24, 49]. In short, for either case, there is a high risk that the synthetic text, once interpreted, is false or misleading. This risk is somewhat mitigated if the source documents are surfaced along with the summary, although user studies would be needed to verify in what contexts and to what extent users click through to check the answers.

Many detrimental use cases of ungrounded generation have been proposed, including LLMs as robo-lawyers (to be used in court [71]), LLMs as psychotherapists [21], LLMs as medical diagnosis machines [40], and LLMs as stand-ins for human subjects in surveys [6, 39]. In all of these cases, a user has a genuine and often life- or livelihood-critical information need. Turning to a system that presents information about word co-occurrence as if it were information about the world could be harmful or even disastrous.

4.5 Summary

In summary, LLMs have several important and often alarming shortcomings as they have been applied as technologies for IA systems. To understand whether and how LLMs could be effectively used in IA, a lot more work needs to be done to understand user needs, IB, and how the affordances both LLMs and other tools respond to these needs. As a step toward beginning that work, we turn to an exploration of the kinds of questions that can be asked in research programs centered not on promoting specific technology but rather on supporting IB.

6.1 Relationships of Trust and Trustworthiness

An ecosystem is a collection of interdependent entities standing in relationship to each other. On the Web, one key type of relationship is that between information providers and information accessors. In this relationship, information accessors desire to find information sources they can trust; information providers desire to show themselves to be trustworthy. Synthetic media break these relationships of trust and trustworthiness, making it harder for people seeking to access information to find sources that are trustworthy—and eventually to be able to trust them even if they have found them.

This disruption of trust began even with simpler forms of synthetic media than LLM-powered chatbots, when search engines switched from providing sets of links to extracting (or abstracting) snippets of text from search results [83] or automatically populating answer boxes [60]. These snippets, even when they are simply extracts from the underlying page, are synthetic media in the sense that the search engine is juxtaposing them to the query and asserting a coherence relation [7] of ‘answer’ between the two. The results can be damaging and dangerous. What follows are four examples, which have since been fixed to varying degrees as a result of having been publicized (both on social media and in the traditional media). (Content warning: the examples beyond the first reflect harmful stereotypes about Indigenous people, Palestinians, and speakers of Kannada.)

In October 2021, Twitter user @soft posted the screencaps in Figure 2 with the comment “The Google search summary vs the actual page.”¹⁰ In this case, it appears that Google extracted the bulleted list, reformatting it as a paragraph and crucially losing the context—the key phrase immediately before the list reading “Do not:”. This took valuable time-sensitive medical advice (a list of things not to do when someone has had a seizure) and turned it into harmful misinformation. (An astute reader might wonder if something is wrong when they hit the parenthetical about tooth or jaw injuries, but we should not expect people in emergency situations to be able to read that carefully or closely.) This case has since been addressed, apparently by the University of Utah removing the list of things not to do from their web page, and pointing to a short video instead.¹¹

Fig. 2.

View Figure

In January 2021, Hank Green noticed and reported on Twitter¹² that Google gave different search snippet results for the closely related queries “when did people come to america” and “when did humans come to america” (Figure 3). Most harmfully, the first of these suggests that ‘people’ in this context refers to white, European people and no one else.¹³

Fig. 3.

View Figure

The examples in Figure 3 show the danger in automatically positioning some text as answer to a query. Figure 4 illustrates another kind of danger: that which arises when IA systems provide answers to questions that are ill formed. Specifically, answering a question serves to validate (accept into the conversational common ground) any presuppositions in the question, through the process called presupposition accommodation [44, 48, 87]. Most utterances have presuppositions; this becomes a problem when the presuppositions are faulty. Consider, for example, trying to answer the question Did you stop smoking? if you have never in fact smoked. Neither Yes nor No will suffice. Instead, you have to directly take issue with the presupposition (I haven’t stopped smoking, because I’ve never started!).

Fig. 4.

View Figure

Until IA systems are able to identify such faulty presuppositions, they run the risk of validating them and thus providing support to bigotry and other kinds of misinformation [46]. Two examples of this are given in Figure 4. In the first,¹⁴ rather than disputing that there is any characteristic headgear of terrorists, Google’s system returns an answer box, complete with photo, about the Palestinian keffiyeh.¹⁵ In the second,¹⁶ rather than disputing the presupposition that there could be such a thing as an ugliest language (in India or globally), Google provided an answer box asserting that the answer is Kannada. Adding insult to injury, the answer box comes with a button offering “Hear this out loud.”

Considering these examples through the lens of trust, we can make several observations. First, in providing snippets and answer boxes, Google is asking users to trust its software not only as a means of finding information sources but as an information source itself. Second, in serving up bigotry, it is betraying that trust. True, Google did not invent the bigotry; the associations driving these results are undoubtedly in underlying corpus. But it is not okay for Google to claim it is just exposing those prejudices, because in these cases it is not. It is reifying them as facts. In thinking about the potential for harm in such reification, value sensitive design asks us to consider different people who might be using the IA system [35, 37]. In this case, we might think of children, who are more likely to be trusting of this kind of information source. What would a Native American or First Nations child make of the result in Figure 3? What would it tell them about how the world at large views them, their people, and their people’s history? This then brings us to the question of trust at the level of communities. To the extent that IA systems reify bigotry as fact, they are the opposite of inclusive and can thus earn the distrust of marginalized communities.

The advent of cheap access to LLMs through APIs has also raised questions of trust for platforms and organizations, and we see such organizations making a variety of choices. For instance, there are platforms like StackOverflow which moved quickly to ban any posting of output from ChatGPT to defend the trustworthiness of the site. StackOverflow’s ban (announced as a temporary policy within days of ChatGPT’s release on November 30, 2022) reads in part (emphasis in original):¹⁷

The primary problem is that while the answers which ChatGPT and other generative AI technologies produce have a high rate of being incorrect, they typically look like the answers might be good and the answers are very easy to produce. There are also many people trying out ChatGPT and other generative AI technologies to create answers, without the expertise or willingness to verify that the answer is correct prior to posting. Because such answers are so easy to produce, a large number of people are posting a lot of answers. The volume of these answers (thousands) and the fact that the answers often require a detailed read by someone with significant subject matter expertise in order to determine that the answer is actually bad has effectively swamped our volunteer-based quality curation infrastructure.

In an accompanying help center post, StackOverflow explicitly calls out trust as a key concept in the decision:¹⁸

Stack Overflow is a community built upon trust. The community trusts that users are submitting answers that reflect what they actually know to be accurate and that they and their peers have the knowledge and skill set to verify and validate those answers.

StackOverflow subsequently reversed this decision, leading to a strike by their volunteer moderators.¹⁹ The strike led to negotiations which ultimately reinstated a fairly strict policy against synthetic content.²⁰

Conversely, according to Futurism, the media website CNET quietly started posting articles generated through a large langaguage model (apparently akin to GPT-3) sometime in November 2022, under the byline “CNET Money Staff,” and this went unnoticed until January 2023.²¹ The articles were purportedly edited and fact checked by human editors but nonetheless contained numerous inaccuracies.²² While, as suggested by The Verge,²³ this move might have brought CNET more traffic and revenue via affiliate ads, it should also have seriously damaged the site’s reputation as a trustworthy news and information source. Unfortunately, CNET is not the only erstwhile news source going down this path. In November 2023, Futurism presented evidence that Sports Illustrated was similarly publishing synthetic content.²⁴ In December 2023, Axel Springer, parent company to Politico and Business Insider, among other publications, announced a partnership with OpenAI that will allow them to “build upon OpenAI’s technology.”²⁵

6.2 Interconnectedness and Synthetic Media Spills

Another facet of the ecosystem metaphor that applies to the Web as an information ecosystem is the notion that structures within the ecosystem are interconnected. Just as a toxic spill into a river can reach its estuary, synthetic media introduced in one part of the Web can contaminate others.

As one example, consider the image on the left in Figure 5, purportedly of a baby peacock (peachick), which has been circulating (without the watermark) on social media. One of us (Bender) encountered it on Facebook on August 4, 2023, and, impressed with how cute it is, shared it.²⁶ Quickly, two of her friends replied flagging the image as fake, and she removed her share of the post. The next day, she attempted to discover what peachicks actually look like, only to find that the synthetic image and others like it were prominent within the Google image search results (Figure 5, right). The synthetic image, rather than staying confined to its source web page, was not only spreading through social media (where users might find it via IF) but also seeping into IR systems.

Fig. 5.

View Figure

The experience reported by the author Jane Friedman in her blog dated August 7, 2023,²⁷ provides another example of how synthetic information can leak through interconnected properties on the Web. Friedman reports that books, likely written with the aid of ‘AI,’ were uploaded to Amazon under her name (presumably through Amazon’s self-publishing facility). While these were not associated with her author page on Amazon, they did end up in her Goodreads profile. The situation has been corrected, likely because Friedman’s social media accounts have broad reach, but Friedman accurately describes how the effort of continually monitoring for and then attempting to address such contamination is an unreasonable burden that falls to authors.

As a final example, we turn to the case of Scots Wikipedia, almost half of which was authored or edited by one American teenager who does not speak Scots.²⁸ Here, the original misinformation was handcrafted and took the form of a misrepresentation of the language itself rather than the information encoded in that language. But automation still plays a role in the contamination of the information ecosystem, even in this case: Wikipedia is a very common source of training data for natural language processing,²⁹ and accordingly, any natural language processing tools for Scots trained on this data are suspect and likely to create further ill-formed Scots text.

6.3 The Value of Provenance and Friction in Maintaining the Information Ecosystem

In this section, we review the ways in which the Web is like an ecosystem and how IA systems can adversely impact that ecosystem, either by originating synthetic media spills or channeling them from their source to other parts of the ecosystem.

In a thread on the social network Mastodon, Rich Felkner elaborates an analogy between efforts to develop ‘AI’ and the fossil fuel industry, writing that both are “[s]eizing and burning something (in this case, the Internet, and more broadly, written-down human knowledge) that was built up over a long time much faster than it could ever be replenished.”³⁰ He elaborates on the value of provenance of information: we can efficiently trust information if we know where it comes from, and we know what sources are trustworthy. His thread continues:

“The ability to produce unlimited amounts of plausible-looking garbage at essentially no cost, and to crowdsource that kind of vandalism to millions of randos by disguising it as something fun, destroys that capability. It’s a DDoS attack on written knowledge.”

Thus, in designing IA systems, we have a responsibility to also think about their impact on the information ecosystem. This means not only refraining from creating tools that leak synthetic media into the ecosystem but also deliberately creating tools that buttress the ability to build trust. In many cases, we believe that trust will result from the productive friction described in Section 5. If we fail to do this, it is not only individuals’ IA that is at stake: when it becomes harder to find trustworthy sources and harder to trust them when we do, we can end up harming societal systems that rely on the public having access to information. These include such consequential domains as public health and democracy.