Controlling institutions with machines
How to enforce policy on a captured but hostile organization
Anti-woke political parties that win elections have a problem: they’re only a few hundred politicians and must act via the civil service. But that civil service is large, left wing and will probably engage in malicious compliance at least some of the time. What to do?
Controlling Institutions With Machines is a series of short essays that lay out concrete, actionable plans for controlling hostile organizations using AI. Designed for leaders in a tough environment where there’s a shortage of loyal staff, this series is a user guide to imposing executive will at scale by leveraging Large Language Models in ways you probably haven’t thought of before.
This essay: introduces the concept and some worked examples.
Efficiency vs control: on the nature of cost reductions or increases in a machine-controlled institution.
Identifying woke employees at scale: How to automatically detect woke employees by exploiting their prediction errors, even when they’re trying to hide.
FRED: Crowdsourcing presidential power: Obtaining extreme levels of grip over the civil service by automating FOIA requests.
Prologue
Imagine you are Trump, and that you win in November 2024. What can you do?
You try to recruit lots of conservatives into the institutions, but that’s difficult. The left is keen on using government power by nature, so it’s easy to recruit them even when pay is low. The right is a broad coalition which includes people who aren’t interested in politics: they just want to grill in peace. For competent but apolitical people, quitting their well-paid private sector job to run some dysfunctional government department is a hard sell.
So, any non-left wing party starts out with an inherent manpower disadvantage. The learned helplessness this creates is a big part of why the UK’s Conservative Party just got thrown out of government. If Trump wins in November 2024 he’ll face this problem again, and it’ll be far worse than before.
A solution can be found in technology. AI allows for radically more automation of management than has ever previously been possible, enabling small numbers of people to swiftly take control of large organizations.
This series of essays shows how it can be done and addresses common questions. We will focus on the civil service, but the approach can be used in any kind of institution.
I. An example system of control
Governments often give away money to left wing lobbyists that are pretending to be politically neutral charities, for protecting birds or whatever. The rules say grantees have to be non-political but the civil service aren’t enforcing it.
You could abolish all charitable grants, but assume for now you’d prefer to spend the political capital elsewhere and would like to fix this more tactically.
You could task a senior civil servant to assemble a committee of more civil servants to review each grant recipient. This won’t work, of course. The senior civil servant will seem obliging and helpful, but when you ask two months later how the work is getting on they’ll inform you that nothing has been done due to some obscure procedural problem they just pulled out of their backside. With sufficient time and fighting a task force will eventually be formed, at which point they’ll conclude that virtually all the grants are going to upstanding pillars of society more neutral than distilled water, except for a few charities who were found to have made a conservative sounding tweet five years ago. Those were of course cut off post-haste, just as ordered.
You spent a lot of effort and time fighting your employees, and they undermined you anyway.
A better approach looks like this. Recruit a small number of trusted allies with programming ability. They write a program that will:
Download a database of all grants being issued to all charities.
For each charity, download the most important pages of the website of that charity, its social media feeds, its press releases etc.
Ask an AI to evaluate each charity’s output against its mission statement, specifically asking it to identify cases of left wing campaigning.
Get another trusted ally to evaluate the performance using some known bad NGOs. It may require a few rounds of tweaking of either the model’s instructions or its parameters, but this is the sort of summarize-and-compare task that LLMs can do well.
The civil service is still involved but their power is severely limited. Once the LLM is properly tuned you mandate that every new grantee is reviewed by it. Staff may file appeals, but they must show the LLM’s assessment doesn’t match the evidence or the rules. Make it clear that those appeals will be reviewed by yourself or a trusted ally and, ideally, that abusive appeals will result in demotions.
Once that process is stable for a few months you can run the program on the full list of grantees to identify every woke NGO, and then defund all of them at once.
The process outlined above:
Is simple.
Requires a few days of work by one trusted ally (or less).
Scales to large workloads, like reviewing every single NGO at once.
Minimizes the risk of malicious compliance as the tasks still handled by humans are precise and mechanical (maintain a database, stop bank transfers, etc).
II. Is AI itself woke?
When ChatGPT was new it claimed it would be better to let a city get nuked than let a black man hear a racial slur. Other models showed similarly manic leftism, so you may be thinking that the idea of using AI to impose anti-woke policy sounds impossible. But as the technology evolved we learned a few things about it.
Firstly, LLMs are extremely malleable. It is very easy to change a model’s behaviours or beliefs via fine tuning, retraining and prompting, especially if you have the weights. There are even ways to disable refusal via direct neural editing. Although LLMs are often released in what the industry calls “safe” (i.e. woke) states, undoing this to create a so-called “uncensored” model hasn’t proven too difficult so far.
Secondly, it seems the manic leftism wasn’t entirely intentional. There’s some evidence that if you train an AI to just mimic the internet it innately becomes left wing. The left write political polemics whilst the right grill smash burgers. This trend was then exacerbated by unwise early practices, like overtraining on the New York Times and academic papers to teach an authoritative and disinterested tone. Luckily, AI companies exist in a free and nearly frictionless market; when Anthropic’s Claude was refusing requests for stupid reasons they lost business to OpenAI and retrained it to be more compliant.
Thirdly, the cost of training and running LLMs is falling very rapidly. Models good enough for the purposes outlined below are now routinely given away for free in editable form. Even if AI companies slide backwards, it’s too late to put the genie back in the bottle.
III. Exposing to non-technical allies
Controlling institutions with AI requires at least a bit of programming. Fortunately that’s a common skill and the sophistication required is low, so I won’t dwell on it here. The task is made easier because AI itself can code quite well now, and there are low-code tools that make using AI with internal data easy.
Still, a group following this plan will focus on recruiting technical staff.
But how does this sort of process look concretely? How do people interact with it? Like, y’know, the actual politicians and civil servants?
The best way to implement it is via a ticketing system. Software like JIRA provides a battle-tested message oriented communication fabric that’s ideal for LLMs. Models can watch tickets and respond as comments are added. Ticket systems also provide buffering, workflow (open → waiting for approval → done), search and editability (so humans can step in and correct errors).
Continuing our example the interface would look like this: for each NGO in the database a new ticket would be created with some initial data like its registered goals and public statements. The LLM would complete an evaluation and add it as a comment to the ticket, simultaneously either closing it (if no problems are found) or moving it to an “attention needed” state. If a human disagrees with the evaluation they have to say why on the ticket and that in turn can automatically escalate to the responsible appointee or ministerial aide.
The key points here are:
The decision itself is made by a machine. The machine can be made into a trustworthy and reliable ally. Humans, not so much.
Humans can override and fix, but are forced to explain.
The user interface is familiar and practical. People already know how to work with ticket systems.
Ticket integrated AI can be deployed within days or weeks of gaining power.
IV. Controlling a media organization
You inherit a state media company and need to make it more trustworthy / less woke (perhaps closing it down entirely is not yet popular enough to consider).
Week 1. Take control of the IT department. It must be run by a trusted ally. Verify that whoever is in charge will execute orders faithfully. Ensure regular two-way communication is flowing and they understand their importance to the mission ahead.
Week 2. Take control of the global content management systems (example) and alter them such that your allies can insert callouts to external APIs that then have veto power over news or content broadcast. Simultaneously, allied programmers start writing scripts to review any submitted content against a series of anti-wokeness policies.
Week 3-8. Start reviewing every piece of content going live, without enforcing the model’s decisions. You and your allies (and only them) should be able to see the AI’s evaluations of each piece of content. It should not be visible to staff what is happening at this time, to avoid distracting fights over unfinished policy. Tune the system’s instructions to ensure accurate classification and steer it towards the desired decisions. This may involve some light database work, e.g. if you want to impose a ‘quota’ of how often a certain topic may appear. The system should fail open; if the AI system can’t do a review within an hour then the content is automatically approved.
Week 9+. Announce to the staff a version of the new policies and explain the moral foundations of the decisions. Announce that enforcement will begin within a week. Do not tell them how their content is being reviewed, only that it’s guaranteed to get reviewed within an hour, that appeals will be read and unsuccessful appeals will be penalized.
Week 10. Enforcement begins. Attention from trusted allies may be needed from time to time to update and adjust the model’s instructions.
If you want to go further, ask the LLM to do some fact checking against sources and also to identify positive examples of neutrally written and fair stories. The latter can be tied to a bonus scheme.
V. More examples
Tee the organization’s email stream into a separate archival, search and classification cluster. The IT department doesn’t need to be involved after that point. Now task LLMs with reading every email and look for attempts to subvert key policies. Use AI to transcribe video conferences to catch cases that aren’t written.
e.g. announce that preferred pronouns may no longer be specified in written communication. Use LLMs to detect cases where simple text match rules are being worked around.
e.g. ban ‘critical theory’ and then let LLMs scan for CT related trainings or events. LLMs will be able to spot cases where people try to disguise their intentions by using euphemisms. Identifications open tickets for allied review.
Fine tune a citation-capable model to make it memorize the full body of law in your country (both statute and case law). Now when you get objections of the form “you can’t do that it’s illegal” you can get a quick double check on the spot without needing an army of allied lawyers. For bonus points set it up to run in the background on your laptop during meetings so you get a live ‘fact check’ of claims being made to your face.
Hook up LLMs to the applicant tracking system and make it review each job advert and hiring packet, looking for DEI hiring criteria.
Note that remote working is beneficial for such a scheme, as it makes it easier for AI to monitor communications whilst simultaneously making it harder for employees to gang up and engage in plots.
The optimal arrangement involves allies setting up companies that are contracted to provide AI review services via direct access to the target organization’s IT systems, as that way the third party can keep the exact LLM taskings secret, making it harder for staff to work around.
The staff will quickly learn both what the spirit of the rules are (due to 100% enforcement) and that they can’t work around them. Those truly committed to their ideology will either quit, be forced underground or be identified and eliminated from the workforce. Those who merely dabbled will test the underlying moral reasoning of the new leadership to see if they can agree with it, and if they can then they will, both to avoid unpleasant cognitive dissonance and because to any civilized person wokeness is clearly repellent. Without pressure to conform a commitment to professionalism and merit will soon re-establish itself.
For more examples and a discussion of AI for efficiency vs AI for control, please see this followup post.
Subscribed after reading this. Interested to see what else you have to say.
I would also add that a lot of government work is just bureaucratic pencil-pushing that could easily be automated. So the activists can not only be monitored via AI but also replaced by it in many cases. Which saves the taxpayers a lot on bloated overhead costs.