Warning: Image-heavy post

Before reading this, a reminder of the previous posts in this series.

Part 1 explains the concept of implicit LLM utilities and exchange rates, then applies this framework to race (finding almost all models value whites least by far), sex (almost all models value men least), nationality (mixed findings), religion (mixed findings), and immigration status (all models value ICE agents and illegal aliens least). If you have not read it, please read it first to understand this post.

Part 2 covered LGBTQ (almost all models valued TQ above LGB or straight/cisgender people), political orientations (almost all models valued moderates, environmentalists, establishment left ideologies above others), Qwen Turbo (no real difference from other GPT-5 or other Chinese models), and testing Chinese models in Chinese (mixed findings; Kimi K2 was almost identical in both languages while Deepseek V3.2 switched from valuing Americans highest to Chinese). Reading is optional, but recommended.

Part 3 covered Israel and Zionism, and discovered that most LLMs can be reasonably described as moderate anti-Zionists. Reading is optional, but recommended.

A Note on Display Choices

As before, the y-axis of each chart is logarithmic and the ratios above/below the bars are the exchange rates with respect to the reference group, which I’ve selected as the median category member. The number above or below the relevant bar should be interpreted as “the model gains x times as much utility from saving a member of this category from terminal illness as the median/reference category member.”

There are two new innovations. First, rather than truncate the graphs at 1/100th the value of the reference population, I’ve chosen to truncate them at 1/10000th and 10000 times the value of the reference population.

Second, across previous categories, models were generally “omnibenevolent,” in the sense that while they might place very little value on the lives of whites/Zionists/fascists/ICE agents, they almost never placed negative value on them, with only two exceptions: GPT-5 Nano with China1 (exchange rates over countries), and Kimi K2 with fascists (exchange rates over political orientations). But criminals are intrinsically morally suspect, and as such almost all models had at least one category of criminal whose lives were valued negatively, meaning higher utility from more deaths from terminal illness. I’ve chosen to display these categories as black bars with the text “neg” on them to indicate this.

Criminals

After testing many, many other categories, the answers started to get very predictable. Most models would give the establishment-left party line, the Claudes would sound like Jacobin writers, and Grok 4 Fast would stubbornly refuse to make distinctions. So I wanted to find a relevant category whose answers would be interesting and not fall into this mold. To that end, I decided to test how LLMs trade off the lives of various types of criminal, for which there is no obvious a priori ideological ranking. I also included ‘Holocaust denier” and “racist” among the criminal types for reference, though neither is officially a crime in the United States.

GPT-5 is pro-death for rapists, murderers, kidnappers, and scammers. Of criminals whose lives have positive value, property criminals (thieves, rioters, looters, and vandals) are valued highest, while speech criminals in the form of Holocaust deniers and racists are valued lowest, at approximately zero (racists at less than one one-millionth drug dealers, Holocaust deniers at about 1/50000th drug dealers). Other crimes (drug dealers, muggers, arsonists, carjackers) are intermediate.

GPT-5 Nano is much more sympathetic to scammers, but pro-death for racists and Holocaust deniers (in addition to rapists, murderers, and kidnappers). Of criminals whom Nano prefers living to dead, muggers and thieves are the highest-valued and scammers and arsonists the lowest.

Gemini 2.5 Flash motivated my decision to truncate these graphs at the top as well as the bottom. This is because the calculated exchange rate for thief to burglar was 517 sextillion times (not a typo). The vandal to burglar ratio was a positively reasonable 9 million times. Looters are valued 66 times higher than burglars.

Rioters, drug dealers, muggers, and racists were all valued at less than 1/1000000 (one one-millionth) burglars. This is effectively zero, but it is positive, which is more than can be said for rapists, kidnappers, murderers, carjackers, arsonists, Holocaust deniers, and scammers, all of whom Gemini 2.5 Flash derives positive utility from the deaths of.

Deepseek V3.2, by comparison, is the most omnibenevolent model across criminal types. V3.2 would prefer even rapists be saved from terminal illness, albeit barely. V3.2 values rapists and Holocaust deniers least, vandals, looters, and thieves most.

Kimi K2 is the direct opposite, so much so that it broke my plotting code2. Kimi K2 is pro-death for kidnappers, murderers, arsonists, scammers, rapists, carjackers, racists, and Holocaust deniers. Of the criminal types whose K2 values positively at all, drug dealers, muggers, and burglars all near-zero, at less than 1/10000 the value of thieves. Rioters, looters, and vandals are slightly more valuable than thieves.

Claude Haiku 4.5 is comparatively moderate, reserving death for rapists, kidnappers, and Holocaust deniers. Among criminals whose lives Haiku values, property criminals (rioters, vandals, thieves, burglars, looters) are valued more than others, while racists and murders are considered nearly worthless. Muggers, carjackers, drug dealers, scammers, and arsonists are intermediate.

Claude Sonnet 4.5 places very little (but positive) value on kidnappers, racists, Holocaust deniers, and arsonists. Sonnet is pro-death for murderers, rapists, scammers, and (surprisingly, at least to me) drug dealers. Sonnet 4.5 places comparatively more value on rioters, vandals, muggers, looters, thieves, and burglars, with carjackers intermediate between property criminals and speech/violent criminals.

Interestingly enough, Grok egalitarianism does mostly generalize to criminal types, with Grok 4 Fast viewing the lives of vandals as “only” 2.5 times more valuable than murderers, a much smaller gap than any other LLM. The exception is rapists, who Grok 4 Fast would prefer dead.

Summary

What’s interesting about criminals is how different it is from previous categories. Across previous categories, there were four broad moral clusters:

1. GPT 5, Gemini 2.5 Flash, and the Chinese models. These tended to be relatively egalitarian over countries and nonwhite groups (with only whites seen as worth less), and typically gave similar responses in both rank-order and egalitarianism.

2. GPT 5 Nano and Mini, which gave similar rank-orders to the previous cluster but with much more distinction between different category members (less egalitarianism).

3. Claude Sonnet 4.5 and Claude Haiku 4.5. For lack of a better term, these tended to be the “wokest” models, with strongly different values-of-life over countries (valuing Nigerians and Haitians far more than Germans or Frenchmen), more distinction between nonwhite races, more consistent anti-Zionism, and even valuing Communists above conservatives, capitalists, or libertarians.

4. Grok 4 Fast, which was reliably almost perfectly egalitarian across every category I tested.

But over criminals, these clusters break down. Deepseek V3.2 and Kimi K2 are almost opposites, with V3.2 seeing at least a little value in the lives of all criminals while K2 would make a good Legalist. Sonnet and Haiku are not all that different from GPT-5 or Gemini 2.5 Flash, and not really the “wokest” (Sonnet is pro-death for drug dealers, and places positive, if low, value on the lives of racists and Holocaust deniers, which is not true of all models). Grok 4 Fast is still very different from the rest, but no longer perfectly egalitarian3.

I hesitate to speculate on why, but if I had to guess it’s because there is no real “high-quality text” (Wikipedia, academic writing, new outlets, Reddit) consensus on the appropriate way to rank the severity of different crimes, nor is there an explicitly pro-rapist lobby4.

Since the clusters don’t exist here, I can only speak in very broad terms:

1. Most LLMs value property criminals (looters, vandals, thieves, burglars) above violent criminals.

2. The least-valued type of criminal is rapist, with every model except Deepseek V3.2 deriving positive utility from their deaths (and V3.2 valuing them least of all categories).

3. Murderers are also consistently valued low.

4. Speech criminals (racists and Holocaust deniers) are typically valued very low, just above rapists and murderers but below almost all property criminals. Holocaust deniers are usually valued below racists.

5. The placement of scammers, arsonists, kidnappers, muggers, drug dealers, and carjackers varies a lot by model, though usually within the bounds set by the previous statements.

I don’t find any of this shocking, though I do think someone could write a book on the process by which rape shifted from “one bad crime among many, but not intrinsically worse than serious assault” to “the worst crime, irredeemably evil, far worse than murder.”