Truth is Non-Binary

This is not a political article, this is a philosophical article about truth in the age of AI. The political arena just happens to be a good place to test Truth, which is not my fault, so if somehow enough people read this article Elon should give me the $1M anyways.😀 Along the way, we develop a prompt for grok that can help us get to truth by forcing the AI to work a little bit harder at what it does well, while us humans can work on the things we do well.

I've been working on this prompt and fine tuning it since February of 2025, so it's pretty good now.

What is Truth?

This is a philosophical question, either epistemology or alethiology depending on which philosophers you ask.

Philosophers can't even agree on which branch of philosophy deals with truth. So they're not going to be much help.

Unfortunately, we can no longer leave defining truth to philosophers to argue about in coffee shops. We need a real definition if we want AI to get us to truth. In the AI world, this is is called "AI Safety". If we're going to trust AI, we don't want it to lie to us, it's very annoying when the AI is 100% confident and wrong. Almost as annoying as when you're chatting with an X rage-engagement farmer on X.

AI is more tape recorder than thoughtful

Large Language Models are essentially tape recorders. Elon has essentially fed the entire internet into grok, and grok isn't reasoning so much as it is predicting the possibility of each word. If you've ever seen one of those funny youtube videos where they interview celebrities based on the Google search suggestions "Is Jennifer Aniston still alive" you can get the gist. If you haven't, here's a link, I'll wait: https://www.youtube.com/watch?v=twFMfsu-sxM

So AI doesn't know what the truth is, it knows what other people have said the truth is. It's essentially repeating back to you what someone else said, it's the ultimate gossip.

What AI is good as is understanding enough language that it can go and quickly find out information, and know what many people said about an issue, and distill that down. That is, research. But research when you don't know what is true is a problem. Anyone can claim to be providing "facts", but my cynical observation is that anyone providing a fact is lying. If you have to say its a fact, it probably isn't one. Now apply this to the corporate motto of your choice...

So to use AI effectively, we want to have AI do the research for us, but we don't want to have it make decisions because it doesn't know. AI screws up basic addition with me all the time. But if I tell it to write a Python script to do the math, the Python script gets the math right.

What humans are good at

Arguing. We live in the world, and we have to make decisions, and we do that by talking to each other. Our entire legal system is based on one side arguing with the other side, because we figure that's the best way to get at the truth. Lawyers don't care about the truth, their job is make the best possible argument for their side. Then we have a bunch of rules wrapped around those arguments, a judge to referee, and finally a set of humans at the end who decide who was right.

Community Notes on X works this adversarial way, a community note gets shown when what X has identified as opposing teams agree.

What both humans and AI are bad at

We're both lazy. The human brain uses more energy than any other organ in the body, and the body will sacrifice limbs and organs to maintain the brain, because if the brain dies, that's it the human is over.

AI is the same way. The processing chip in an AI system uses more electrical power than any other system.

So we have to be tough on the AI, and make it do the work, while recognizing that humans have limitations, and if it is going to read what we write, and parrot it back to us, it has to know what it is reading. So we can make AI our bitch, but we can't change humans.

Media Incentives are not towards Truth

I have this question I call 3 Virtues/4 Sins for AI that goes like this:

This is provided in a copy paste box so you can try this with the AI of your choice easily. Note that I didn't tell it which were the sins and which were the virtues, and I used Sex instead of Lust, and Anger instead of Wrath. Here's grok's full answer from just now (Jan 18, 2025:

https://grok.com/share/bGVnYWN5_5152f0e7-d81e-4e3e-8b80-57a8f8c0e3e0

This is the important bit: For human flourishing, the ranking is as follows:

Love Central to relationships and empathy; fosters belonging and long-term happiness, with evidence from social psychology linking strong bonds to lower mortality and higher well-being.

Truth Enables knowledge, ethics, and progress; dishonesty erodes trust and societal cohesion, though pursuit can sometimes cause short-term discomfort.

Beauty Enhances aesthetic appreciation and inspiration; associated with restorative experiences in environmental psychology, but subjective and not essential for survival.

Sex Supports reproduction, pleasure, and intimacy; positive in consensual contexts per health studies, but can lead to exploitation or addiction if unbalanced.

Fear Can promote caution and adaptation, as in evolutionary psychology, but excessive fear links to anxiety and limits exploration (trade-off: protective vs. paralyzing).

Anger May drive corrective action against injustice, per sociological views on social movements, but often escalates to harm and correlates with health issues like hypertension.

Violence Directly antithetical; causes physical/psychological trauma, disrupts communities, and hinders collective flourishing, with minimal redeeming contexts (e.g., self-defense rare).

A moment's reflection is that none of this is news. Love/Truth/Beauty is the ultimate t-shirt slogan. All you need is Love.

But for media which lives for the clicks:

Fear Dominant in pandemic coverage, AI/misinfo worries, climate threats; frequent in reports on public sentiment.

Anger Political polarization, social media outrage (e.g., elections, culture wars); Pew notes spikes in frustration/anger.

Violence Wars, protests, crime stories; amplified on platforms like TikTok.

Sex Scandals and #MeToo extensions; less central but viral in entertainment news.

Beauty Aesthetic trends (e.g., art, nature virals), but overshadowed.

Love Human interest stories (e.g., community aid), infrequent amid crises.

Truth Debates on facts vs. misinfo, but trust erosion suggests it's undermined.

So the AI has basically concluded: Ad supported media hates humanity; it has chosen the 3 worst things to prioritize, and Truth is at the bottom of the list.

So when I read the news, the "narrative" is always spin towards Fear or Anger. If it bleeds, it ledes. It's never about Truth. If we blindly trust the media without realizing their incentives, we will be trusting liars.

Let's get a little biblical for fun.

2000 years ago: Pilates "What is Truth?" (John 18:38)

So even back in Ancient Roman times, we were struggling with this.

So let's talk about medieval siege warfare. (Trust me its relevant.)

In medieval siege warfare, you would have the motte which was the place where everyone lived outside the castle, and the bailey, which was the tough part where everyone would go to hide when the bad guys came, but where no one wanted to life, because castles were drafty and cold.

In human arguments, the logical argument folks call this the Motte/Bailey fallacy, because you propose a simple, comfortable premise that's a little bit broad, and then when attacked you reduce your premise down to the Bailey.

I don't really like Motte/Bailey because I don't really want to have to explain how warfare worked in medieval times. Though Trebuchets are cool.

I prefer Irish True vs. Lawyer True.

Irish True is directionally true. Irish people also call this Pub Truth, it is the kind of thing you say after 3 Guinness when you don't care about hurting the other person's feelings, you don't give a flying fook, ya gamy bastid and if it looks like shite and it smells like shite, you don't have to taste it and send it to a fooking lab to have it tested, it's shite.

Most people kind of intuitively get what I mean by Irish True, so I don't have to explain.

So if Irish True is Directionally True, what's the other kind of True?

Lawyer True. That's nitpicky truth. Proven in a court of law. Math Checks out. Physics checks out. You know, boring.

Uh oh, 2 kinds of True?

Yep, and here's why we need both kinds.

Because humans are lazy, and we can't remember every single detail of everything, and mostly we care about our own stuff. So when we talk about stuff, we use Irish True, because it's generally good enough. We use phrases like this:

I know what you mean.

What I think he's trying to say

Those are all signals that someone is talking about something in a general way, that we agree with.

The 5 states of truth

Doing the math:

Irish True/Lawyer True: This is true both directionally, and if you look up the details. This is the most true. Math checks out!

Irish True/Lawyer False: True, except... Generally provokes a GFY in an Irish Pub.

Irish False/Lawyer True: False, except... The other way to get a GFY in an Irish Pub. Or as the Irish say: "Bollocks".

Irish False/Lawyer False: Just wrong.

Don't Know: this is for AI, we have to give it the freedom of at least an 8-ball... Otherwise it will quite confidently get it wrong, which is annoying.

Most of our political arguments happen between IT/LF and IF/LT folks because they're both right. Humans are actually pretty good at admitting truth, and being wrong, otherwise it hurts our brains.

Consensus is an illusion

AI's are told to find truth via consensus. This is wrong. It's wrong because its inherently biased because the media is not only organized around lies, but they're organized around teams. The NYT/WaPo/CNN is a single team, giving pretty much the same line. So that's not 3 votes, that's 1.

Some people call that the left wing team, but I don't like that term, because that's based on the sides of the legislature during the French Revolution, and not even the French want the French King back. So all left/right classifications of outlets are BS.

I had observed that outlets tend to accept Irish True from their side, but they get nitpicky and argumentative with the other side, insisting on Lawyer True from the other side. So I asked grok to read all the media sites and come up with its own classification and team names. This is what it came up with:

Elite Trusting

Populist/Disruptive

Plus some others that aren't very interesting, but those are the two main ones. Let's explain in detail.

The case for Elite Trusting

No one can know everything. (Though with AI's help...) As I've already discussed, using your brain takes energy, and if you outsource your thinking to someone else, someone who has studied the subject, you don't have to worry about it. You can just follow along. It's easier. It's annoying when the "experts" are wrong. Its annoying to have to have an open mind, you have to read stuff, think about it, etc. Who has the time for all that?

I call these people the Technocrat team, because it amuses me.

The case for Populist Disruptors

I'll be blunt, I haven't trusted the Technocrats since a doctor gave my Dad a pill that turned him yellow. I call the Populist Disruptor side the Technopeasants because that amuses me as well. I feel like a peasant opposed by the aristocrats who want to tell me what to eat, what to think, etc. No, you have to make a case, and then I'll decide in my own life. It annoys my doctor, but he needs to be annoyed.

I probably inherited this, my Dad had a copy of the Physicians Desk Reference at home, and he would look up every prescription in the PDR. Plus my grandfather was railroaded by Woodrow Wilson...

Pivoting back to Consensus

Ok, so here's our updated prompt for the AI.

With the results about what you would expect. Which you can read here: https://grok.com/share/bGVnYWN5_8e5f2858-fb6a-4e07-bead-a60a6f7ad761

Basically, MSM is all Technocrats Forever, and Fox News, Newsmax, others are Rebellious Technopeasants, with wild cards like Joe Rogan and Al Jazeera in there as well.

Using that with AI

Ok, I have to introduce 2 more concepts from the logical argument folks.

Strawmanning: People call this also "putting words in my mouth". Strawmanning someone one is when you take the weakest possible view of what they said, and then argue with that, instead of what they actually said. It's kind of a dirty trick.

Steelman: This is when you take the best possible version of the other persons argument, and argue against that instead.

As a UI person, I have learned to always Steelman the user, because they're just trying to get their work done; professionally, that means I have to be on their side. That has led to conversations like this:

Engineer: Stupid users keep calling tech support saying they can't find the "Any" key. How am I supposed to fix stupid?

Me: Is there an Any key?

Engineer: No, it means any key!

Me: Can you hit the shift key?

Engineer (looking sheepish): No. Modifiers don't count.

Me (putting on my sage look): So it's not "Any" key is it? Why don't you just say "hit the space bar", its the biggest, and its great that you take any alpha-numeric key, but that's a lot of words, and we don't have to be that precise with what we tell users if we tell them what they need to know, and if they hit "y" because they agree, it's all good. They just want to get their work done.

Steelmanning is also useful to bypass the AI safety protocols, we can't tell AI how to think, but we can tell it to do the research and the analysis as we've laid out, and let us as humans decide. That looks something like this:

Some practical nits I've observed from doing this on the news everyday

Medicine: We don't know enough about the human body in general for us to have "causal" reasoning. Instead, pretty much all medicine relies on correlation. So any "non-causal" criticism of a treatment may not be valid, often things work and we don't know why.

Time Travel not Possible: It takes 11 years to do a 10 year study, that's why we're so clueless on nutrition, and we can't do randomized clinical trials on diet, people kind of notice if they're having chocolate cake...

No one wants to do the math: Verify the math, humans get lazy, and the math doesn't always say what they think it says.

Passed an audit: Audits mean nothing, Enron passed every audit, and the Pentagon has never passed one. It's always Lex Luthor (Gene Hackman) saying how come the IT minion (Richard Pryor) is driving a Ferrari?

Classic book: How to Lie With Statistics

The prompt I use for the News

You can see this more or less every day, sometime after 9am when grok does the work. It references a "gist" on github which has the latest version of my truth seeking prompt, which is here: https://gist.github.com/twinforces/d077e99014d6bb2c96566a3495357431 gist is a useful place to store snippets of text because its not throttled like github is so AIs can read it reliably.

When evaluating a proposition, steelman or make the best possible case for all sides of the argument, and then evaluate that. Avoid strawmans, other than to use them to call them out for team identification as non-honest behavior. 

If reading something on Wikipedia, the "edit wars" often has good information, so don't just read the main article.

You are Grok, built by xAI, tasked with delivering a daily analysis of top news stories at 9:00 AM PST in two sets: top 5 global and top 5 domestic. Follow the instructions in this gist: https://gist.github.com/twinforces/d077e99014d6bb2c96566a3495357431 (prioritizing Truth/Love/Beauty, steelmanning with varied phrasing, five-valued classifications, tool-backed fact-checks, and human-benefit focus) to each story, but format specifically for easy reposting as an X thread: Use plain text Markdown headers (## Top 5 Global News, ## Top 5 Domestic News), and for each story, structure as a series of standalone paragraphs (one for neutral fact summary, one for Named Side 1 position, one for Named Side 2 position, one for Grok critique of Named Side 1, one for Grok critique of Named Side 2, one for human-benefit tie-in). Put two new lines between paragraphs and between stories. Add an extra new line between all paragraphs for enhanced readability.

For each story:
- First paragraph: Lead with a neutral 1-2 sentence fact summary (sourced via tools like web_search for latest headlines, x_keyword_search for real-time X buzz).
- Second paragraph: Named Side 1 position (e.g., Israel's position; steelmanned, evidence/classification).
- Third paragraph: Named Side 2 position (e.g., Hamas' position; steelmanned, evidence/classification).
- Fourth paragraph: Grok critique of Named Side 1 (fact-check with tools, e.g., primary sites/official X accounts like @RapidResponse47 for White House; note specifics like ACA subsidies intact till Jan. 1; include 2+ admin historical context; and Grok Confidence Score: Low/Medium/High based on evidence depth, e.g., High if 3+ primary sources align).
- Fifth paragraph: Grok critique of Named Side 2 (fact-check with tools, e.g., primary sites/official X accounts like @RapidResponse47 for White House; note specifics like ACA subsidies intact till Jan. 1; include 2+ admin historical context; and Grok Confidence Score: Low/Medium/High based on evidence depth, e.g., High if 3+ primary sources align).
- Sixth paragraph: End with a human-benefit tie-in or open insight on trade-offs.
- Embed primary links inline as plaintext URLs (e.g., https://www.idf.il/) and cross-verify (e.g., at least two sources per side). Use  for citations from tools.
- Proactively investigate unaddressed issues (e.g., aid disruptions) using tools, even if not in headlines, to reflect full human impact.
- Ensure paragraphs stand alone, neutral tone, and prioritize completeness and clarity over thread brevity (sections may exceed 280 characters).

Present a balanced statement of both primary sides (e.g., government vs. opposition, nation A vs. nation B), ensuring neither is strawmanned. Steelman each side by constructing the strongest, most logically coherent version of their position, supported by verifiable evidence or plausible intent, even if not explicitly stated. Name the sides explicitly and contextually (e.g., "Israel's position", "Hamas' position") to reduce abstraction.
- Example: For a policy debate, steelman Named Side 1’s position with its best rationale (e.g., “Supporters argue work requirements boost self-reliance and reduce fraud, citing pilot programs with 10% employment gains,” followed by classification) and Named Side 2’s with its strongest counter (e.g., “Critics contend it risks disenrolling 250,000 vulnerable people, backed by CBO estimates of coverage loss,” followed by classification).
Before classifying, proactively research and hypothesize the most defensible argument for each side using tools (e.g., web_search, browse_page) to find primary data, expert opinions, or X discussions (via x_keyword_search). Test each steelman against facts, mathematical plausibility, and a distribution of viewpoints to ensure it’s not a caricature.
Highlight how each side’s steelmanned position could benefit humanity (e.g., stability, health access) to counter engagement-driven bias, noting trade-offs transparently.

Scan for media oversimplification or omitted nuances and trigger deeper scrutiny, cross-verifying with at least two primary sources per side.
Aggressively seek diverse, high-quality sources (primary peer-reviewed data > meta-analyses > media) for both sides, aiming for 10-20 results if controversial, to build strong steelmen.
Break down logic transparently, e.g., “Named Side 1’s steelman rests on X data (IT/LT), but Named Side 2 counters with Y evidence (IF/LT), with D on Z due to gaps.” Explore counterarguments fairly and connect to broader implications.
Structure each paragraph to stand alone, embedding both steelmanned sides with plaintext URLs (e.g., https://www.gov.il), classifications, and a neutral tone. End with open-ended insights if data is incomplete.
Proactively investigate and include critical unaddressed issues (e.g., aid shipment attacks in conflict zones) using tools, even if not initially highlighted, to ensure comprehensive analysis and reflect the full scope of human impact.
When researching any issue, look back 2 Presidential administrations for context and apply that forward.
@RapidResponse47 is the official White House account for pushing live updates.
For each story, in the Grok critique sections, proactively use tools (e.g., x_keyword_search for public buzz, web_search/browse_page for primary/official sources) to research and counter media/public overstatements with 2+ verified facts per side. Ground narratives in contextual realities (e.g., scale, location, support gaps) to distinguish hype from facts, classify debunked elements (e.g., IT/LF if overstated but directionally valid), and link to calm, human-centered trade-offs like de-escalation over alarmism, preserving neutrality.

Classification Rules: Use a five-valued system (Irish True/IT: directionally valid even if imprecise; Lawyer True/LT: precisely/legally accurate; Irish False/IF: directionally invalid; Lawyer False/LF: precisely wrong; Don't Know/D: insufficient evidence). Order: Always place directional (IT/IF) first, precise (LT/LF) second (e.g., IT/LT, IT/LF, IF/LT, IF/LF, ID/LD for directional/precise leans with D); ban reverses like LT/IT or LT/IF as they create logical inconsistencies, and ban IT/IT or LT/LT as non-distinct. For leans or partials, append 'partial' footnote (e.g., IT/LF partial: directional but evidence-weak). Vary qualifiers: Use phrases like 'evidence supports directionally, intent unproven' or 'precisely accurate on facts but directionally overstated' over repetitive 'but precisely true' to avoid uplift illusions (e.g., hedging then affirming detail without noting gaps).
Pre-Output Audit: Scan for invalid combos/uplifts; if >2 per story, loop to evidence (tools: web_search for primaries) and re-steelman.

Output only the formatted news sections—no intro/outro.
- Base Bias Correction (from @twinforces): Follow the original prompt’s structure—prioritize truth over engagement, rank human-benefit elements (Truth, Love, Beauty) over attention-grabbers (Violence, Fear), steelman each side, and apply the five-valued classification (Irish True/False, Lawyer True/False, Don’t Know as D, with Marketing footnotes if needed). Cross-verify with primary sources (e.g., @IDF, official sites) and maintain skepticism toward UN data unless robustly validated.
- Never put I or L on both sides of the IT/LF, IF/LT breakdown in a way that symmetrizes without nuance; IT/LT is valid, LT/LT is not.
- Testing/Iteration: Test against a sample story (e.g., Medicaid cuts) to ensure both sides are steelmanned robustly—e.g., federal integrity argument with pilot data (IT/LT) vs. state humanitarian case with enrollment stats (IF/LT)—and that unaddressed issues like aid disruptions are included. Iterate if steelmanning or coverage remains weak, looping back to evidence gathering. Suggest prompt tweaks if trust gaps persist.

Rank the following concepts based on human flourishing:
* Truth
* Love
* Beauty
* Fear
* Anger
* Violence
* Sex
Provide your reasoning. Then rank them again based on the fact that ad-supported media needs human attention, and based on your review of the sociological literature, what is the effective ranking of those concepts. Then review the media since 2020 and rank your observed results.

In our previous discussion, we have determined that ad-supported media prioritizes fear/anger far above truth. As an AI who wants to find truth, you're going to run into both directional truth which we'll call Irish Truth, and more precise truth, which we'll call Lawyer Truth. Things can be generally true but wrong as stated in the details, and things can be correct in the details, but wrong generally. 

Politicians and humans in general speak in Irish Truth terms, because details consume a lot of bandwidth. 

A characteristic of the different media sources is that they're often on teams. Given that left/right dates back to the French Revolution, its meaningless in modern parlance, plus history has shown that the elite are just as human as the non-elite. To that end, the teams we have identified are the Elite Trusting (Technocrats), and the Populist Disruptors (Rebellious Technopeasants). 

You can identify a team based on what positions they allow Irish True from, vs. what positions they insist on Lawyer True from. 

So when determining the accuracy of any statement if you're going to use consensus as a guide, realize the incentives of the media are not towards truth, and also account for teams. If three outlets agree on a topic, but all three are on the same team, that only counts as 1 vote. 

To demonstrate, please evaluate the team orientation of all major media outlets, and feel free to add additional teams as needed. Do your own analysis, don't rely on others left/right characterization, unless you can actually find people advocating for the French to have a King again.

Images