Can our systems be adapted to AI challenges? Or rather human manipulation of such systems of information? Would current models suffice to capture added “information”? Probably, but policy process studies are in their infancies.
At the end of the day, the whole point is that we may have to update our theories to account for these changes, and maybe in a way we don't like that much : empirically *and* normatively. Maybe information *should* originate in humans, and theories should acknowledge that. Human agency over machines
And on the output side, the Dutch government literally resigned over an algorithm nobody could fully account for (illegible decisions). It seems policymakers and citizens alike are caught in a feedback loop where problems are related to machines.
www.politico.eu/article/dutc...
On the input side, "it is impossible to tell whether survey respondents are real or bots." In this case, how priors "transmit" to policymakers ? If we can't certify that, how do we know priors reaching policymakers are genuine at all? Illegible priors bsky.app/profile/josh...
Priors are stubborn. But how do "priors transmit" collectively without reliable information? Without any human certification of what is real? And from an epistemological standpoint: isn't the whole point of the policy process to respond to priors through reliable, human-grounded information?
On the input side: signals can now be manufactured quite easily. On the output side: decisions are increasingly delegated to algorithms actors themselves don't understand (the Dutch childcare case, Parcoursup in France, risk-scoring systems ; in short, incresing system/screen-level bureaucracies)
Our theories (and policy processes) assume we can trace the causal chain: a problem is signaled; institutions process it; a decision is made. Theoretical coherence (and legitimacy of the PP) both depend on that traceability. That assumption won't hold much longer to me.
Thank you @alongcamejones.bsky.social! I'd push further though. The deeper issue may not be about power dynamics, but about the whole policy process legibility (for citizens and policymakers alike).
#publicpolicy #policytheories #policyprocess #policyprocesstheories #agendasetting #LLMs #AI #epistemology #politicalscience @alongcamejones.bsky.social @paulcairney.bsky.social @cmweible.bsky.social @lr-bg.bsky.social @sam-workman.bsky.social @eakoebele.bsky.social @raulpachecovega.bsky.social
I would be genuinely glad to engage in an honest disciplinary conversation on this. I think it is time we stop looking away. In my mother tongue, we have a saying: "in the land of the blind, the one-eyed man is king." Let us make sure that proverb does not become our discipline's defining motto. 7/7
Produced by a company called CiviClick. Policy decisions are already beginning to respond to machines rather than to humans. In response, in my article I propose a new framework, or at least new theoretical boundaries, designed to strengthen the epistemological foundations of our theories. 6/7
By every conventional standard, this decision should be a textbook case of public policy responding to democratic signals. There was only one problem. Those comments were manufactured, produced by LLMs. 5/7 www.latimes.com/environment/...
In June 2025, a California air quality board voted 7 to 5 to reject anti-pollution rules that would have prevented approximately 2,500 premature deaths. The board had received over 20,000 citizen comments opposing the proposal. 4/7
If we are truly honest, can we still claim our theories are valid?
In my view, if we do not rebuild our theories for the age of machines, they will die. In a new article available as a preprint, I use agenda-setting as a case study to show that our theories are already dying. 3/7
Not a day goes by without a new story about machines replacing what was previously the exclusive province of humans: writing and producing information. Yet public policy theories were built in a world where information production was a fundamentally human affair. That world no longer exists. 2/7
Policy theories are facing an epistemological crisis. One would have to be blind not to see each day LLMs swallowing up our public sphere, that space where we are supposed to reason together and build a shared deliberative common ground: the foundation of policymaking. 1/7 🧵
doi.org/10.31235/osf...
Comment des #méthodes d’analyse textuelle peuvent-elles nous aider à mieux comprendre le discours de #démission de F. #Legault ?
Je publie dans @ledevoir.com une analyse disponible ici de façon interactive : antoinelemor.github.io/assets/prese...
www.ledevoir.com/opinion/idee...
3/5 ⚖️ 6 justifications for resigning vs 99 defending his record
🏛️ Dominant theme: party politics (24 mentions), ahead of healthcare (15) and economy (13). He talks more about his party than public policy.
2/5 📈 72% of the speech dedicated to his legacy
This isn't a resignation speech. It's a legacy speech. Legault explains less than he defends.
On tone: 51% pragmatic + confident, only 7% alarmist. Combative peaks (12%) are targeted and reserved for defense and sovereignty issues.
7/9
On policy content: focus on diplomatic method (24%) rather than concrete measures (3%). The "how" of negotiating matters more than the "what".
6/9
On France's positioning: 85% agency score. France presented as active agent (46%), partner (23%), leader (8%). Victim in only 3% of cases.
5/9
On sentiment toward key actors: explicit European anchoring. France (+0.67) and Europe (+0.51) positive. US slightly negative (-0.06), China clearly negative (-0.42). Ukraine and India = preferred partners (+0.75). Russia? Barely mentioned.
4/9
On worldview: Geopolitical Anxiety Index of +0.08. Neither catastrophist nor triumphalist. 61% of geopolitical frames emphasize opportunities, 39% threats.
3/9
On rhetorical structure: deliberate alternation between confident and alarmist registers. Almost symmetrical. Macron systematically balances threats and opportunities. A (very) controlled rhetorical pendulum. 2/9
Full analysis, slides, code and methodology:
antoinelemor.github.io/blog/2026/nl...
Anyone can use and critique. That's science. 9/9
Tools used (all open source):
→ Transcribe-tool (audio transcription):
github.com/antoinelemor/Transcribe-tool
→ LLM_Tool (annotation framework):
github.com/antoinelemor/LLM_Tool 8/9
On responses to journalists: 72% partial answers.
Concrete questions (governance, diplomacy, legality) remain unanswered.
An operation without an exit plan? 7/9
antoinelemor.github.io/blog/2026/nl...