Pandora's Hope
Pandora's Hope: Essays on the Reality of Science Studies (Harvard University Press, 1999) collects Latour's major essays on the philosophical status of ANT's realism, written during the height of the 'science wars' of the 1990s. The book defends the position that scientific knowledge is genuinely about reality — not merely social construction — while insisting that the access to reality is mediated by heterogeneous networks of humans, instruments, institutions, and texts that cannot be reduced to either the 'social' or the 'natural' side of the modern constitution. For the AI moment, the book's analysis of technical mediation is particularly important: the essay 'On Technical Mediation' traces how tools transform the actors that use them, providing the philosophical foundation for treating Claude as a mediator rather than an intermediary.
The Machinery of Consensus — Contrarian ^ Opus
There is a parallel reading that begins from the political economy of scientific production rather than its philosophical status. Latour's account of technical mediation and circulating reference elegantly describes how knowledge moves through networks, but it remains curiously silent about who owns the networks, who decides which chains of reference get built, and whose realities get stabilized as facts. The person-with-gun hybrid may redistribute agency philosophically, but the gun manufacturer, the lobbyist, and the legislator remain untouched by this redistribution. When we apply this to AI, the silence becomes deafening: the human-with-Claude hybrid obscures the fact that Anthropic decides what Claude can and cannot say, what values it embodies, what knowledge it recognizes as legitimate.
The circulating reference that produces scientific facts requires massive infrastructure — laboratories, journals, funding agencies, tenure committees. Each node in this chain filters what can pass through according to criteria that have less to do with reality-access and more to do with institutional survival, career advancement, and the maintenance of existing power structures. Climate science provides the textbook case: decades of circulating reference establishing anthropogenic warming, yet the political-economic networks that determine action remain unmoved. For AI, this means that however sophisticated our philosophical account of human-AI mediation becomes, the actual capabilities, biases, and purposes of these systems will be determined by the companies that build them, the investors who fund them, and the regulators who constrain them. The hybrid of human-with-Claude may produce novel capabilities, but those capabilities serve interests that Latour's framework renders invisible. Technical mediation becomes technical mystification when it fails to ask: mediation on whose behalf?
In the AI Story
The book was written in the aftermath of the Sokal hoax and the broader 'science wars' that pitted scientific realists against scholars in science studies, cultural studies, and critical theory. Latour was a particular target of the realists, who accused him of dissolving scientific facts into social construction. His reply, developed across the essays, was that neither 'construction' nor 'fact' named what actually happens in scientific practice. Facts are constructed — that is, produced through concrete chains of laboratory work, instrumentation, and institutional stabilization — and they are also real, in the sense that they refer to features of the world that do not reduce to the social processes that produce the knowledge.
The book's most influential essay for technology theory is 'On Technical Mediation,' which develops a four-part analysis of how tools modify the actors that use them. Latour traces the gun-control debate — 'guns don't kill people, people kill people' versus 'guns kill people' — to argue that both sides are wrong. Neither the gun alone nor the person alone explains what happens when a gun is fired. What explains is the hybrid: the person-with-gun, an actor-network whose capabilities, intentions, and consequences cannot be attributed to either party in isolation. The analysis generalizes to every technology, including AI.
For AI, the implications are direct. Claude does not passively transmit the human user's intention (the 'tool' view). Claude does not determine the outcomes independently of the human (the 'autonomous AI' view). What produces outcomes is the hybrid: the human-with-Claude, an actor-network whose capabilities are jointly constituted. The amplifier metaphor maps onto the failed 'guns don't kill people' position; it attributes all agency to the human and treats the technology as passive. The accurate description is hybrid: outcomes emerge from the combined action of human and AI within a network that includes both and neither alone.
The book also introduces the concept of 'circulating reference' — the chains of translation through which scientific claims connect to the phenomena they describe. A geological claim about the forests of the Amazon is supported by a chain that runs through soil samples, color charts, field notes, published diagrams, and expert testimony. No single link in the chain has direct access to 'nature'; but the chain as a whole produces reliable knowledge about the forest. The same structure applies to AI outputs: no single moment in Claude's processing provides direct access to the truth, but the chain of training, prompting, verification, and correction can produce reliable results — provided the chain is maintained.
Origin
The book collected essays Latour had written throughout the 1990s in response to the science wars and to ongoing philosophical debates about realism, construction, and the authority of science. It was published by Harvard in 1999 and became one of his most accessible works, frequently used in graduate seminars on science studies.
Its reception was significant in repositioning Latour as a defender of scientific realism against the more radical versions of social constructionism with which he had sometimes been conflated. The repositioning did not fully convince his critics, but it established clearer philosophical ground for ANT than had been available in the earlier laboratory studies.
Key Ideas
Non-reductive realism. Scientific knowledge is about reality, but the access to reality is mediated by heterogeneous networks that cannot be reduced to either 'the social' or 'the natural.'
Technical mediation. Tools transform the actors that use them. The hybrid — actor-with-tool — is the proper unit of analysis, not actor or tool alone.
The gun essay. Gun-control debates are a paradigm case for analyzing how agency is distributed in person-tool assemblages. Neither 'guns don't kill people' nor 'guns kill people' captures the hybrid reality.
Circulating reference. Scientific claims connect to the world through chains of translation. No single link provides direct access; the chain as a whole produces reliability.
Factishes. Latour's portmanteau for entities that are simultaneously fabricated and real — facts whose constructedness does not diminish their truth-claims.
Appears in the Orange Pill Cycle
Networks Within Power Structures — Arbitrator ^ Opus
The proper frame for technical mediation requires both Latour's philosophical precision and the contrarian's political-economic grounding. When we ask "How does technology transform action?" Latour's hybrid analysis is essentially correct (95%) — the person-with-gun or human-with-Claude genuinely creates emergent capabilities that belong to neither component alone. The philosophical move from attribution to distribution of agency captures something fundamental about how tools work. But when we ask "Whose interests do these hybrids serve?" the contrarian view dominates (80%) — the ownership and governance of technical systems shapes their mediating effects in ways that Latour's framework undersystematizes.
The circulating reference concept illuminates how knowledge achieves reliability through chains of translation, and this remains valid (70%) for understanding how AI systems can produce truth-effects despite lacking direct world-access. Yet the contrarian correctly identifies (75%) that these chains pass through institutional filters that shape what counts as legitimate reference. Climate science exemplifies both sides: the chains of reference genuinely connect to atmospheric reality, but their political efficacy depends on power structures that ANT's symmetrical treatment obscures. The factish concept — entities both constructed and real — offers genuine philosophical insight, but needs supplementation with analysis of who controls the construction process.
The synthesis requires a doubled framework: technical mediation operates within fields of power. The human-with-Claude hybrid does create novel capabilities through genuine mediation, not mere tool-use or determination. But this mediation occurs within an infrastructure owned by Anthropic, trained on data selected according to specific criteria, and governed by policies that encode particular values. The proper unit of analysis is neither the hybrid alone nor the power structure alone, but the hybrid-within-structure — technical mediation as it actually occurs, shaped simultaneously by the logic of tools and the logic of capital.
Further reading
- Bruno Latour, Pandora's Hope: Essays on the Reality of Science Studies (Harvard University Press, 1999)
- Bruno Latour, 'On Technical Mediation' in Common Knowledge 3 (1994)
- Bruno Latour, We Have Never Been Modern (Harvard University Press, 1993)
- Graham Harman, Prince of Networks (re.press, 2009)
- Isabelle Stengers, Power and Invention (University of Minnesota Press, 1997)