The disclosed internal research identified design interventions that would reduce the documented harms. On the Institute's reading of the public record, these were not implemented in effective form.
The internal documents at Facebook did not merely document harm. It identified specific design interventions that would reduce it. This distinction matters because it transforms the evidentiary record from a story about ignorance into a story about decision-making. The company did not learn that its platform harmed adolescents and then face the difficult question of what to do about it. The company learned that its platform was associated with harm to adolescents, and the disclosed research pointed to specific design interventions that could reduce it. The Institute assembles these into a remediation catalog and pairs each with the engagement/revenue tradeoff its attention-inventory model implies. This catalog-with-tradeoffs is the Institute's synthesis of what the disclosed findings support — not a costed catalog attributed to Facebook's own researchers; the public record does not show per-intervention revenue projections produced internally and evaluated against each other.
The first intervention category involved like count visibility. Internal researchers documented that the public display of like counts was a primary driver of social comparison among adolescent users. The mechanism was direct: a visible like count converts a social interaction into a quantified ranking. The recommendation was to hide like counts, or at minimum to hide them for users under eighteen. The projected welfare benefit was measurable: reduced social comparison, reduced anxiety around posting, reduced fixation on quantified social validation. The projected revenue cost was also measurable: reduced engagement with like-based feedback loops, reduced posting frequency driven by like-count anxiety, reduced time-on-platform associated with like-count monitoring.
The second category involved content recommendation algorithms. Internal research documented that the Explore page and algorithmic feed surfaced appearance-related, diet-related, and body-comparison content to adolescent users at rates that correlated with increased body dissatisfaction and negative self-evaluation. The recommendation was to modify the algorithmic ranking for users under eighteen to reduce exposure to content categories that the research had specifically identified as harm vectors. The projected welfare benefit was reduction in the documented body image effects. The projected revenue cost was reduction in engagement generated by the highest-engagement content categories for adolescent users — which were, precisely, the categories the research linked to the harm.
The third category involved direct content filtering for diet and body image material. The research recommended limiting adolescent exposure to content promoting extreme dieting, body transformation, and appearance comparison. This was not a general content moderation recommendation. It was specific to content categories that the internal research had identified as associated with measurable harm in the adolescent population. The implementation would have required content classification at scale, which the company already performed for advertising purposes but had not applied to adolescent welfare.
The fourth category involved usage time prompts and friction mechanisms. The research recommended implementing prompts that would alert adolescent users to time spent on the platform, encourage breaks, and introduce friction into the infinite scroll architecture. The projected welfare benefit was reduction in the cumulative exposure effects that the research had documented: the relationship between time-on-platform and the severity of body image, comparison, and mood effects was dose-dependent, and reducing the dose would reduce the effect. The projected revenue cost was, by definition, the reduction in time-on-platform — which is the primary metric against which advertising revenue is calculated.
The underlying research was the company's own — conducted by its employees and presented through its internal channels before the Haugen disclosure. The remediation catalog and the per-intervention revenue tradeoffs above are the Institute's synthesis of what that research supports; they are not presented as a costed catalog that the company's own researchers are known to have assembled and ranked internally.
The organizational structure in which these recommendations were evaluated is the key to understanding why they were not implemented. The structure is not complex. It does not require allegations of malice or conspiracy. It requires only the recognition that Facebook, like all publicly traded companies operating on an attention-inventory business model, evaluates product decisions against engagement and revenue metrics.
The attention-inventory model operates on a specific economic logic: the company sells advertising impressions. The value of those impressions is determined by the number of users, the time each user spends on the platform, and the precision of the targeting. Product changes that increase any of these three variables increase revenue. Product changes that decrease any of them decrease revenue. This is not a distortion of the business model. It is the business model.
The welfare research documented that the design features producing adolescent harm were the same features producing adolescent engagement. The like count that drives social comparison also drives posting frequency and return visits. The algorithmic amplification of appearance content that produces body dissatisfaction also produces the highest engagement rates among adolescent users. The infinite scroll architecture that enables cumulative exposure also maximizes time-on-platform. The harm and the revenue are not independent variables. They are two outputs of the same design choices, and they are inversely related: reducing the harm requires reducing the engagement, and reducing the engagement reduces the revenue.
This is the Revenue-Welfare Inversion. It is not a failure of corporate ethics. It is an incentive structure. When the organizational decision-making process evaluates product changes against revenue impact, and when the welfare-improving changes carry revenue costs, the predictable output of the decision-making process is that the welfare-improving changes will not be implemented. The inversion does not require anyone to decide to harm adolescents. It requires only that no one in the decision chain has the authority or incentive to accept the revenue cost of not harming them.
The AE series (Saga VIII) documents this incentive architecture in detail as a general feature of attention-economy business models. The Instagram case is a specific instance of the general pattern. The internal research provided the company with the information necessary to modify the architecture. The incentive structure ensured that the information would not produce the modification. The research existed. The remediation catalog existed. The Revenue-Welfare Inversion determined the outcome.
The company's response to its internal research was not total inaction. It was selective action — and the pattern of selection reveals the operating logic.
Like count hiding was tested. In 2019, Instagram began testing hidden like counts in several countries, including Canada, Australia, Brazil, and Ireland. The feature removed the public display of like counts on posts, allowing only the post's author to see the number. In 2021, the option to hide like counts was made available to all users globally — as an opt-in feature. The default remained visible like counts. The distinction between a default-on and an opt-in implementation is not trivial. Default settings determine the experience of the vast majority of users because the vast majority of users do not change defaults. An opt-in feature that addresses the documented harm but is not enabled by default addresses the public relations problem without modifying the actual user experience at scale. The feature exists. The harm architecture persists.
Algorithmic modifications for adolescent users were not implemented in the period between the internal research findings and the Haugen disclosure. The Explore page continued to surface content through the same engagement-optimized ranking for users under eighteen as for all other users. The content categories that internal research had specifically identified as harm vectors — appearance-focused content, diet content, body comparison content — continued to be algorithmically amplified for adolescent users because they continued to generate high engagement rates among adolescent users. The algorithm was not modified. The harm mechanism continued to operate as documented.
Content filtering for diet and body image material was minimal. The company's content moderation policies addressed content that violated community standards — graphic self-harm imagery, content promoting eating disorders in explicit terms — but did not address the broader category of appearance and diet content that the internal research had identified as the harm vector. The distinction is between content that is explicitly harmful in its messaging (pro-anorexia communities, self-harm tutorials) and content that is harmful by virtue of its structural position in the comparison architecture (aspirational body imagery, fitness transformation content, diet culture). The former was moderated. The latter was amplified.
Usage time prompts were implemented — and designed to be dismissed with a single tap. Instagram introduced a "Take a Break" feature (rolled out in December 2021, after the Haugen disclosure) that reminded users after a specified time period. The implementation included a prompt that could be dismissed instantly, no default-on time limit, and no friction mechanism that would interrupt the scroll architecture. The feature created the appearance of a time-management tool without modifying the underlying design that the research had identified as the engagement mechanism. The infinite scroll continued. The autoplay continued. The notification architecture continued. The usage prompt existed alongside the engagement architecture rather than modifying it — a speed limit sign posted on a road with no enforcement.
"Facebook has made numerous changes to protect teen users: restricting DMs from adults, implementing time management tools, hiding like counts in some markets. The claim that nothing was done is factually incorrect."
The distinction is between safety features and architectural modifications. Restricting direct messages from adults to minors addresses a specific safety risk: predatory contact. It does not modify the engagement architecture that the internal research identified as the harm vector. Implementing time management tools addresses the appearance of concern about usage duration. It does not modify the infinite scroll, autoplay, notification, and algorithmic amplification systems that the internal research documented as the mechanisms through which duration produces harm. Hiding like counts in some markets as an opt-in feature addresses the public relations dimension of the like-count research. It does not change the default experience for the population the research identified as harmed.
The pattern is consistent: changes that address specific safety risks without revenue impact were implemented. Changes that address the core engagement architecture — the algorithmic amplification, the comparison engine, the reward loop — with revenue impact were not. Safety features and architectural modifications are different categories. On the public record, the former were implemented; the latter were not. This is consistent with the Revenue-Welfare Inversion, not a refutation of it.
How internal welfare research moves within a company's organizational structure is itself a design decision with structural consequences. On the Institute's reading, the disclosed record is consistent with welfare research being processed through legal/liability review rather than product review — and the Institute models that routing as determining the organizational response. (This legal-versus-product routing is the Institute's analytical frame, not a documented routing instruction from Meta; Haugen's testimony characterizes the failure as resolving profit-versus-safety conflicts in favor of profit.)
Legal review and product review are different institutional functions with different outputs. Product review assesses a finding in terms of design modification: given this research, what should the product do differently? The output of product review is a design change — a modification to the algorithm, the interface, the default settings, the content ranking. Product review converts research into product decisions.
Legal review assesses a finding in terms of liability exposure: given this research, what is the company's legal risk? The output of legal review is a legal strategy — privilege assertions, document retention policies, public statement drafting, regulatory positioning. Legal review converts research into liability management. It does not produce product changes because product changes are not within its institutional function.
On this model, processing welfare research through legal rather than product channels converts a welfare finding into a liability assessment, and the organizational consequence follows from the routing. Legal review does not have the authority to order product modifications. It does not evaluate design alternatives against welfare outcomes. It evaluates documents against litigation risk. Research processed primarily through that function produces one category of output: legal strategy. On the Institute's reading, the design modifications the research pointed to were not evaluated by the institutional function with the authority to implement them.
This is the architecture of non-decision the Institute models. On this reading, the company did not decide not to implement the remediation catalog so much as route it to an institutional function that could not implement it. The non-decision the Institute describes is structural: no individual executive need have rejected the welfare recommendations; an architecture that directs them to a function whose output is legal strategy rather than product design would have the same effect. On this reading, the remediation catalog was processed as a liability matter rather than emerging as a product specification.
The routing the Institute models would itself have been a decision. On this reading, some organizational protocol — rather than any single executive — would have processed research documenting harm to adolescent users through legal channels rather than product channels. On the Institute's argument, that routing is the decision that matters. Everything that followed was the predictable output of the routing architecture. Once the research entered the legal function, the only possible organizational outputs were legal outputs: privilege designation, litigation preparation, regulatory positioning, public relations strategy. The product remained unchanged because the product function never received the research through channels that would have produced product changes.
The pattern documented in SG-001 through SG-004 is not unique to Facebook. It is a structural pattern that recurs whenever an industry generates internal research documenting harm linked to its own products, and when remediation of that harm would reduce the revenue generated by those products. The tobacco industry provides the clearest historical parallel — and the parallel is precise, not metaphorical.
Tobacco companies' internal research programs, beginning in the 1950s and continuing through the 1990s, produced findings documenting the health effects of smoking. These findings were not limited to harm documentation. They included identification of harm-reduction measures: filter design modifications that would reduce tar and particulate delivery, nicotine reduction strategies, additive modifications that would reduce the formation of specific carcinogens. The remediation catalog existed. It was evaluated. It was not implemented.
The reason it was not implemented follows the same structural logic as the Instagram case. Implementing harm-reduction measures implied acknowledging that the product was harmful. Acknowledging that the product was harmful created liability. The organizational incentive structure therefore disfavored both the acknowledgment and the remediation. The optimal strategy under this incentive structure was to suppress the research, delay the remediation, and publicly contest the evidence of harm — which is what the industry did for approximately four decades.
The organizational logic is identical in both cases: once internal research documents harm, any remediation based on that research functions as an implicit acknowledgment of the harm. The acknowledgment creates liability exposure. The liability exposure exceeds the welfare benefit of the remediation in the organization's decision calculus — because the organization's decision calculus weights financial outcomes, and the liability cost of acknowledging harm exceeds the financial benefit of reducing it. The remediation is therefore foregone not because the organization does not know how to implement it, but because implementing it is structurally adverse to the organization's interests as defined by its incentive architecture.
The TB series (Saga VII) documents this pattern in detail for the tobacco industry. The Instagram case extends the pattern to the attention economy. The specific products are different. The specific harms are different. The organizational logic — the incentive structure that converts internal welfare research into a liability problem rather than a design problem — is the same. The Foregone Remediation is not a coincidence. It is a structural feature of industries in which harm arises from the same product features that generate revenue.
The evidentiary significance of the Foregone Remediation is specific, and it is distinct from the significance of the internal research itself.
The internal research documented in SG-001 through SG-003 speaks to institutional knowledge: the company's own research documented that its platform was associated with measurable harm in adolescent users. This finding is significant but not, by itself, sufficient to establish the full evidentiary case. A company might know that its product produces harm and be unable to modify the product without destroying its utility. A company might know that its product produces harm and be genuinely uncertain about which modifications would reduce it. A company might know that its product produces harm and be in the process of developing and testing remediation.
The Foregone Remediation eliminates each of these defenses. The company was not unable to modify the product — its own researchers identified specific, technically feasible modifications. The company was not uncertain about which modifications would work — its own researchers projected the welfare impact of each. The company was not in the process of developing remediation — on the Institute's reading of the public record, the identified modifications that carried revenue costs were not implemented in effective form.
What the Foregone Remediation establishes, on the Institute's reading, is a connection between institutional knowledge and institutional inaction. The company's own research documented the harm, and pointed toward feasible remediations. Yet the harm-reducing changes the research supported were not implemented. The Institute models this outcome as the product of an organizational structure — welfare research processed through legal rather than product channels, product changes evaluated against revenue metrics, and opt-in rather than default-on implementation of the modifications that were partially deployed — that, intentionally or emergently, tends to ensure that knowledge of harm does not produce remediation of harm.
The Foregone Remediation is the evidentiary link that transforms the Instagram case from a story about unforeseen consequences into a story about organizational architecture. The harm was foreseen. The remediation was identified. The architecture prevented the remediation. The architecture was not modified. The harm continued. Each of these statements is documented in the company's own internal record. The Foregone Remediation is what the Institute reads the company's own evidence to document about the company's own decisions.
This is the structural core of the Instagram case as an evidentiary matter. Not what the company did not know — it knew. Not what the company could not have done — it could have. Not what the company intended — intention is irrelevant to the structural analysis. What the company's organizational architecture was designed to produce: the continuation of the harm-generating design in the presence of the harm-documenting research in the absence of the harm-reducing remediation. The Foregone Remediation is not an accusation. It is a description of how the system worked.
Internal: This paper is part of The Instagram Files (SG series), Saga IX. It draws on and contributes to the argument documented across 22 investigations in 5 series.
External references for this paper are in development. The Institute’s reference program is adding formal academic citations across the corpus. Priority papers (P0/P1) have complete references sections.