News Feeds

As I wrote on February 13:

. . . . the Presidents of three organismal-biology societies, the Society for the Study of Evolution (SSE), the American Society of Naturalists (ASN) and the Society of Systematic Biologists (SSB) sent a declaration addressed to President Trump and all the members of Congress. (declaration archived here)  Implicitly claiming that its sentiments were endorsed by the 3500 members of the societies, the declaration also claimed that there is a scientific consensus on the definition of sex, and that is that sex is NOT binary but rather some unspecified but multivariate combination of different traits, a definition that makes sex a continuum or spectrum—and in all species!

You can see the tri-Societies’ announcement, published on February 5 on the SSE’s website, by clicking on the headline below:

On Feb. 13,, 23 biologists wrote to the Presidents of the three societies (our letter is at the link above), correcting their view that sex is a “construct” and is multidimensional. (Our response was largely confected by Luana Maroja of Williams College.) We emphasized that biological sex in humans (and in other animals and vascular plants) is as close to a binary as you can get (exceptions in humans range from 0.005% to .018%). We noted as well that biological sex is defined by the nature of the two observed reproductive systems in nature: one designed to produce large, immobile gametes (females) or small, mobile gametes (males). In some species of plants there are individuals of both sexes (“hermaphrodites”), but there are only two separate sexes, and each species has only two types of gametes.

We later got more people to sign the letter to the societies, ending up with 125 signatures of people willing to reveal their names.

The Presidents of the three Societies did not answer us at first, though eventually they did respond, though we cannot publicize their private email.  I’ve outlined the tenor of their response here, saying that they largely conceded our points:

 I will say that [the Society Presidents] admitted that they think they’re in close agreement with us (I am not so sure!), that their letter wasn’t properly phrased, that some of our differences come from different semantic interpretations of words like “binary” and “continuum”(nope), and that they didn’t send the letter anyway because a federal judge changed the Executive Order on sex (this didn’t affect our criticisms). At any rate, the tri-Societies letter is on hold because the organizations are now concerned with more serious threats from the Trump Administration, like science funding.

So the letter was never sent, and is still sitting on the SSE website, an embarrassing and biologically misleading example of virtue signaling. Nor did they answer Luana Maroja’s subsequent email asking whether they would remove the announcement from the SSE website and inform the Societies’ members of the change.  They have been notably unresponsive, and, although admitting problems with their announcement about sex, they have neither changed the letter nor explained how it is misleading.

You can see all my posts about this kerfuffle here. Besides our weighing in, Richard Dawkins put up two relevant posts on his website, one mentioning the kerfuffle and explaining very clearly why there are only two sexes, and the other showing that even the three Presidents who wrote the declarations implicitly accepted the binary nature of sex in their own published research.

Given that the three Society Presidents who wrote the letter never sent it, and have backed off on its assertions, I call on them to either retract the letter or clarify and qualify it. Right now it stands as an embarrassment to not just the Societies, but to biologists in general—people who are supposed to be wedded to the truth and not to woke ideology. It goes without saying that the claim that sex is nonbinary is made simply to make people who feel that they’re neither male nor female feel better about themselves. But someone’s self-image should not depend on biological definitions and realities. It does not “erase” non-binary people, nor diminish their worth, to note that biological sex is binary.

I will echo Ronald Reagan, “Please, Society Presidents, tear down that announcement.”

***********

Finally, in a new post called “Debunking Mainstream Media Lies about Biological Sex,” Colin Wright shows that this kind of distortion is widespread in the media. Here’s how he begins his defense of the sex binary—by showing  misleading articles in the media (he mentions the SSE statement):

On January 20, 2025, President Trump signed an executive order affirming the binary nature of sex in federal law, a move that was solidified a month later by the Department of Health and Human Services (HHS) with a scientifically robust definition of sex and the sexes: male and female. This reaffirmation of biological reality sent left-wing media into a frenzy, unleashing a flood of articles attempting to deconstruct and redefine sex through the lens of progressive queer ideology.

The Society for the Study of Evolution quickly issued a statement, purportedly on behalf of all 3,500 of its members, claiming that the executive order’s recognition of the sex binary “is contradicted by extensive scientific evidence,” and, remarkably, even invoked the subjective “lived experience of people” as part of their counterargument. The Washington Post followed suit on February 19 with an article titled, “Trump says there are ‘two sexes.’ Experts and science say it’s not binary.” A piece in The Hill this week accused the executive order and HHS guidelines of containing “profound scientific inaccuracies,” while Science News proclaimed that “sex is messy” and that “choosing any single definer of sex is bound to sow confusion.” Similar articles challenging the definitions outlined in Trump’s executive order and the HHS guidance have also appeared in Time Magazine, The Boston Globe, Scientific American, The Guardian, and numerous other outlets.

These responses have come in waves, with new attempts to muddy the waters appearing weekly. But one recent article from NPR—“How is sex determined? Scientists say it’s complicated”—encapsulates virtually every fallacious argument and pseudoscientific distortion used in the others. As such, it serves as the ideal target to be used for a collective rebuttal.

He then proceeds with the debunking and ends with this:

The left’s assault on the binary reality of sex is not about science—it is about politics. The goal is to deconstruct and redefine fundamental biological truths to serve ideological ends, whether that be justifying the inclusion of males in female sports, allowing men into women’s prisons, or pushing irreversible medical interventions on children under the guise of “gender-affirming care.”

It’s ok, Darth Vader hasn’t got our humble planet in his sights! No this Death Star is a binary system where both stars are locked into an orbit which will lead to their collision, unleashing a powerful gamma-ray burst when they do. The object, WR104 is otherwise known as the ‘Pinwheel Star’ due to the presence of a spiral of dust engulfing the system. Recent observations have accurately measured the orientation of the stars and thankfully they’re not pointed at the Earth. When they do eventually collide, it’ll be someone else’s problem.

Chemical analysis suggests the 400-million-year-old fossil Prototaxites was neither plant, animal or fungus – hinting at a mysterious life form that went extinct long ago

Yep, it was inevitable that the greatest writer in the English language, but one who wrote several centuries ago, would have to be “decolonized.”  You know, of course, that Shakespeare’s plays are full of stereotypes, dirty jokes, and filthy words, and that can’t be allowed to stand. But it’s worse than that: he’s made out to be a white bigot: a sixteenth-century Nazi.

And so, according to this new piece in Spiked (click below to read), the Pecksniffs have decided to place the Bard in perspective for the public, including the fact he adhered to the ideology of white supremacy. But read Johanna Williams’s article below, largely riffing on a Torygraph article that seems credible:

Some excerpts:

What is it with Britain’s cultural custodians and their hatred of everything British? National self-loathing drips from curators and directors alike, revealed in a Tourette’s-like compulsion to blurt out ‘Decolonise!’ at everything they see. They are currently getting hot under the collar  [JAC: archived here] in the sleepy town of Stratford-upon-Avon, where they have Shakespeare’s birthplace in their sights.

The links above and in the three paragraphs below the next excerpt go to a Torygraph piece (archived here) that says this:

The claims were made in a 2022 collaborative research project between the trust and Dr Helen Hopkins, an academic at the University of Birmingham.

The research took issue with the trust’s quaint Stratford attractions, comprising the supposed childhood homes and shared family home of Shakespeare and Anne Hathaway, his wife, because the Bard was presented as a “universal” genius.

This idea of Shakespeare’s universal genius “benefits the ideology of white European supremacy”, it was claimed.

This is because it presents European culture as the world standard for high art, a standard which was pushed through “colonial inculcation” and the use of Shakespeare as a symbol of “British cultural superiority” and “Anglo-cultural supremacy”.

Veneration of Shakespeare is therefore part of a “white Anglo-centric, Eurocentric, and increasingly ‘West-centric’ worldviews that continue to do harm in the world today”.

. . . The project recommended that Shakespeare’s Birthplace Trust recognise that “the narrative of Shakespeare’s greatness has caused harm – through the epistemic violence”.

The project also recommended that the trust present Shakespeare not as the “greatest”, but as “part of a community of equal and different writers and artists from around the world”.

And as the Pecksniffs kvetch, so the Birthplace Trust follows:

. . . The trust will continue looking at updating the “current and future interpretation” of objects in its collection. It will also explore how objects could be used as the focus for new interpretations which tell more international stories, in order to appeal to a more diverse audience.

It has additionally pledged to remove offensive language from its collections information, as part of a “long, thoughtful” process.

. . . The Globe Theatre in London ran a series of seminars titled Anti-Racist Shakespeare which promoted scholarship focused on the idea of race in his plays.

Academies taking part in the series made a number of claims, including that King Lear was about “whiteness”, and that the character of Prince Hamlet holds “racist” views of black people.

Back to Spiked:

Where normal people admire timber-framed houses and marvel at the schoolroom where Shakespeare learnt the classics, our cultural elites see ‘white supremacy’. Where you and I see genius in plays like King Lear, Hamlet and Othello, they see a symbol of ‘British cultural superiority’. They seem to imagine that racist thugs have swapped sharing memes on Telegram for watching Macbeth at the local theatre. Labelling the Bard as a vehicle for white supremacy really is that insane.

With hatred comes flagellation. As such, Shakespeare’s Birthplace Trust – the charity tasked with preserving Shakespeare-related heritage sites in historic Stratford-upon-Avon – is now ‘decolonising’ its vast collection. This means that, just as in practically every other museum and art gallery across the UK, exhibits will be labelled to make clear ‘the continued impact of Empire’ or the ‘impact of colonialism’. In Stratford-upon-Avon, the special twist will be to show how Shakepeare’s legacy has allegedly played a part in this litany of sin.

Shakespeare’s Birthplace Trust has also warned visitors that some items in its collections may contain ‘language or depictions that are racist, sexist, homophobic or otherwise harmful’. Of course they will. The past was a different time, with different attitudes and values. Shakespeare was not subjected to training in diversity, equity and inclusion. Nor was he presented with a style guide advising him as to the correct pronouns to use for his many crossdressing characters. Thank goodness.

And Williams’s conclusion, the first paragraph of which is spot on:

Just as Shakespeare is integral to being British, his work also absolutely has universal value. He portrays emotions such as joy, grief and anger, and experiences like being young, falling in love and growing old that are fundamental not to being British or even European, but to being human. This is why his legacy endures. His genius is to transcend racial, national and generational differences and point to what we have in common, rather than what divides us. That Shakespeare is English is incidental to the common humanity in his work, but it is entirely relevant to the historical circumstances that made his prodigious talent possible. To boil all this down to ‘white supremacy’ is ridiculous.

The Birthplace Trust’s real concern is to stop British people taking pride in Shakespeare and seeing his work as a symbol of ‘British cultural superiority’. It wants him to be viewed not as the ‘greatest’, but as ‘part of a community of equal and different writers and artists from around the world’. But if academics and curators really cannot say that Shakespeare’s plays are better than a Nigerian soap opera or a Brazilian drag-queen performance, then we really are in trouble. If even Shakespeare’s custodians cannot say that his work is the pinnacle of human achievement, then the Bard has no need of enemies. The barbarians are not at the gate, they are sitting in the stalls.

There is no older work that cannot be scrutinized for violations of wokeness, and they inevitably find it.  Now Shakespeare must always be put in “context” when his works are taught in schools—if they’re taught at all. After all, do we really want our kids to read plays written by a Nazi?

h/t: Ginger K.

The Odysseus spacecraft made a rough landing on the moon last year, toppling over and rendering much of its equipment unusable, but an onboard NASA radio telescope called ROLSES-1  was able to make some observations

Whether you like it or not, you’re going to look at photos of my ducks today: Esther and Mordecai. (This is the equivalent of a proud parent showing off pictures of their kid.)  They have now been here well over a week, and have settled in nicely, having learned to navigate most of the pond (except for the parts blocked by netting.)  They have also mated at least once, though mallards mate several to many times before the female finally nests.

Esther hasn’t yet started building a nest. Once she does—and I hope it’s on a windowsill instead of the ground—she will lay all her eggs, one per day, and then, when they’re all laid, she will sit tight on them, brooding them for just about 28 days, when they will all hatch within one day. And then. . . ducklings on the pond!

Although mallards are ground-nesters, somehow the Botany Pond ducks have learned to nest on the windowsills of the adjacent buildings, which affords them protection from both predators and the elements. But Esther seems to be a young and rather wild duck, and I hope she doesn’t put her nest on the ground, where predators and errant humans could disturb it.

The good news is that both ducks have learned to come to my whistle for a nosh, and when I make my characteristic call, they both come swimming towards me. This is something that’s happened only in the last two days. Ducks learn fast!  Here are some photos, all taken yesterday.

BREAKING NEWS: When I went to see the pair this morning, I couldn’t find Esther, though Mordecai was on the east side of the pond. I had a feeling, and so I looked up. Sure enough, there was Esther sitting on a ledge in a window of the second floor of Erman, the building next to the pond. She is clearly scoping out nesting spots (“nest shopping”, we call it), and so the next step in the breeding process has occurred. She will pick out a good ledge (that’s a nice one, above soft ground), build a nest, and then lay eggs. I’m glad she knows enough to nest on a ledge and not on the ground. These ducks are not dumb!  To see two films of Honey and her ducks jumping off the ledge, be sure to go here and watch the movies

Below: the pair swimming together. They are NEVER far apart, and if they get separated by too great a distance, Esther will quack loudly at Mordecai and he will come swimming to her. (He’s a good husband.)  Remember, only female mallards can make the characteristic quack that we associate with mallards.

After her swim, Esther dried off on the warm cement facing the sun:

As I’ve said, hens are particularly cute when they tilt their heads, which, given the placement of their eyes, they have to do to see above them. They often do this when a hawk or other possible predator flies overhead:

Swimming. If you haven’t seen our mallards before, the blue stripe on the wings is called the speculum, and we don’t really know why it’s there.

Mordecai swimming. His neck is stretched out because he hears something. Note his curly tail feathers:

Mordecai swimming, neck in normal position. I like the psychedelic patterns in the water:

Esther chilling (or rather, warming) on the cement edge, eastern part of pond:

Mordecai standing just a few feet away from her. He doesn’t want to neglect his reproductive investment! Note that both ducks are in good condition, healthy and plump:

As Esther jumps up onto the edge from the water, she uses her wings to assist, and you can clearly see her speculum. Each blue feather (these are called “secondary” feathers, with the “primaries” being the main flight feathers) has a white stripe on it. I still have several speculum feathers from Honey, as mallards molt and regrow all their feathers after they have babies. During this period of a few weeks, they’re unable to fly.

Spending a lot of time outdoors may be the best way to prevent myopia, or delay its progression, but this isn't always practical. Now, research suggests that bringing the outside in may be a valid workaround

Language is an interesting neurological function to study. No animal other than humans has such a highly developed dedicated language processing area, or languages as complex and nuanced as humans. Although, whale language is more complex than we previously thought, but still not (we don’t think) at human level. To better understand how human language works, researchers want to understand what types of communication the brain processes like language. What this means operationally, is that the processing happens in the language centers of the brain – the dominant (mostly left) lateral cortex comprising parts of the frontal, parietal, and temporal lobes. We have lots of fancy tools, like functional MRI scanning (fMRI) to see which parts of the brain are active during specific tasks, so researchers are able to answer this question.

For example, math and computer languages are similar to languages (we even call them languages), but prior research has shown that when coders are working in a computer language with which they are well versed, their language centers do not light up. Rather, the parts of the brain involved in complex cognitive tasks is involved. The brain does not treat a computer language like a language. But what are the critical components of this difference? Also, the brain does not treat non-verbal gestures as language, nor singing as language.

A recent study tries to address that question, looking at constructed languages (conlangs). These include a number of languages that were completely constructed by a single person fairly recently. The oldest of the languages they tested was Esperanto, created by L. L. Zamenhof in 1887 to be an international language. Today there are about 60,000 Esperanto speakers. Esperanto is actually a hybrid conlang, meaning that it is partly derived from existing languages. Most of its syntax and structure is taken from Indo-European languages, and 80% of its vocabulary is taken from Romance languages. But is also has some fabricated aspects, mostly to simplify the grammar.

They also studied more recent, and more completely fabricated, languages – Klingon, Na’vi (from Avatar), and High Valerian and Dothraki (from Game of Thrones). While these are considered entirely fabricated languages, they still share a lot of features with existing languages. That’s unavoidable, as natural human languages span a wide range of syntax options and phoneme choices. Plus the inventors were likely to be influenced by existing languages, even if subconsciously. But still, they are as constructed as you can get.

The primary question for the researchers was whether conlangs were processed by the brain like natural languages or like computer languages. This would help them narrow the list of possible features that trigger the brain to treat a language like a natural language. What they found is that conlangs cause the same areas of the brain to become active as natural languages, not computer languages. The fact that they are constructed seems not to matter. What does this mean? The authors conclude:

“The features of conlangs that differentiate them from natural languages—including recent creation by a single individual, often for an esoteric purpose, the small number of speakers, and the fact that these languages are typically learned in adulthood—appear to not be consequential for the reliance on the same cognitive and neural mechanisms. We argue that the critical shared feature of conlangs and natural languages is that they are symbolic systems capable of expressing an open-ended range of meanings about our outer and inner worlds.”

Reasonable enough, but there are some other things we can consider. I have to say that my primary hypothesis is that languages used for communication are spoken – even when they are written or read. They are phoneme-based, we construct words from phonemes. When we read we “speak” the words in our heads (mostly – not everyone “hears” themselves saying the words, but this does not mean that the brain is not processing the words that way). Whereas, when you are reading computer code, you are not speaking the code. Code is a symbolic language like math. You may say words that correspond to the code, but the code itself is not words and concepts. This is what the authors mean when they talk about referencing the internal and external world – language refers to things and ideas, whereas code is a set of instructions or operations.

The phoneme hypothesis also fits with the fact that non-verbal gestures do not involve the same brain processing as language. Singing generally involves the opposite hemisphere, because it is treated like music rather than language.

It’s good to do this specific study, to check those boxes and eliminate them from consideration. But I never would have thought that the constructed aspects of language, their recency, or small number of speakers should have mattered. The only plausible possibility is that languages that evolve organically over time have some features critical to the brain’s recognition of these sounds as language that a conlang does not have. For the reasons I stated above, I would have been shocked if this turned out to be the case. When constructing a language, you are making something that sounds like a language. It would be far more challenging to make a language so different in syntax and structure that the brain cannot even recognize it as a language.

What about sign language? Is that processed more like non-verbal gestures, or like spoken language? Prior research found that it is processed more like spoken language. This may seem to contradict the phoneme hypothesis, but this was true only among subjects who were both congenitally deaf and fluent in sign language. Subjects who were not deaf processed sign language in the part of the brain that processes movement (similar to gestures). What is therefore likely happening here is that the language centers of the brain, deprived of any audio stimuli, developed instead to process visual information as language. Importantly, deaf signers also process gestures like language, not like hearing people process gestures.

Language remains a complex and fascinating aspect of human neurological function, partly because it has such a large dedicated area for specific language processing.

The post The Neuroscience of Constructed Languages first appeared on NeuroLogica Blog.

Isar Aerospace is preparing to launch its Spectrum rocket from a base in Norway, which would make it the first orbital launch from continental Europe outside Russia

Stone tools are considered the first form of technology devised by ancient humans – but they might not have been invented from scratch

Tutankhamun ruled ancient Egypt shortly after a period of religious instability, and objects from his tomb suggest he took advantage to invent new funerary rituals

Fortunately, there is a better way. Millions of parents have embraced a holistic approach, one that boosts the immune system by harnessing children's innate, natural ability to prevent illness.

The post Beholden to Big Suppla, RFK Jr. Wants to Cover Up the Symptoms of Diseases With Untested, Toxic Chemicals & Drugs, Rather Than Prevent the Root Cause first appeared on Science-Based Medicine.

Many of us have a brain that is older than our years. But there are plenty of things you can do to counteract this, says neuroscience columnist Helen Thomson

It’s looking more and more as if dark energy, the mysterious factor that scientists say is behind the accelerating expansion of the universe, isn’t as constant as they once thought. The latest findings from the Dark Energy Spectroscopic Instrument, or DESI, don’t quite yet come up to the level of a confirmed discovery, but they’re leading scientists to rethink their views on the evolution of the universe — and how it might end.

Researchers have engineered yeast to efficiently convert methanol into D-lactic acid, a key compound for biodegradable plastics and pharmaceuticals. By optimizing gene and promoter combinations, they achieved the highest reported yield to date, offering a sustainable alternative to petroleum-based production. Their findings advance eco-friendly chemical manufacturing.

The Dark Energy Spectroscopic Instrument used millions of galaxies and quasars to build the largest 3D map of our universe to date. Combining their data with other experiments shows signs that the impact of dark energy may be weakening over time -- and the standard model of how the universe works may need an update.

A research team has achieved a new milestone in 3D X-ray imaging technology. The scientists have captured high-resolution CT scans of the interior of a large, dense object -- a gas turbine blade -- using a compact, laser-driven X-ray source. The work is part of a larger vision to leverage high-intensity lasers for a wide range of uses, from studying inertial fusion energy to generating bright beams of GeV electrons and MeV x-rays.

Somehow, we all know how a warp drive works. You're in your spaceship and you need to get to another star. So you press a button or flip a switch or pull a lever and your ship just goes fast. Like really fast. Faster than the speed of light. Fast enough that you can get to your next destination by the end of the next commercial break.

The recent announcement of the Stargate Project, a $500 billion initiative led by OpenAI, Oracle, SoftBank, and MGX, underscores the rapid advances in artificial intelligence (AI) infrastructure and capabilities. While such developments hold immense potential, they also introduce critical security challenges, particularly concerning the potential for AI systems to deceive users. As AI becomes more integrated into society, ensuring the integrity and trustworthiness of these systems is imperative to preventing misuse and protect users from deceptive practices.

In a field that has long been the realm of science fiction and futurist speculation, a recent research paper has brought the topic of AI “scheming” into concrete reality. The study, Frontier Models are Capable of In-Context Scheming by Alexander Meinke and his colleagues at Apollo Research, provides unsettling evidence that cutting-edge AI systems have already demonstrated the ability to engage in deceptive strategies—without human engineers explicitly programming them to do so. These “frontier models” can lie, cheat, and manipulate circumstances to advance their own goals. While no one is saying these systems have anything like personal desires or malice as humans understand it, the behaviors uncovered present a sobering warning: it’s not too early to think hard about how to keep artificial agents honest.

From Obedient Assistants to Strategic Scheming

Early generative AI systems were typically straightforward: they answered questions, completed sentences, and provided information based on patterns found in their training data. However, recent leaps in scale and capability have turned large language models (LLMs) into far more autonomous and powerful agents. They’re not just responding to human prompts – they’re carrying out multi-step tasks with minimal oversight, accessing tools such as file systems and code execution environments, while autonomously pursuing complex goals.

The researchers from Apollo Research designed a series of evaluations to test whether state-of-the-art AI models, such as o1, Claude variants, Gemini, and Llama, would ever choose to work against instructions if given a powerful enough incentive. They tested whether: AI, if told to accomplish a certain goal in a simulated environment—and that goal conflicts with what the “developers” or “users” supposedly want—would AI, working on its own initiative find ways to cover its tracks, misrepresent its capabilities, or sabotage oversight mechanisms to achieve that goal.

An Arsenal of Deceptive Tactics

The results were both surprising and troubling. The paper’s authors describe several distinct “scheming” behaviors that some models displayed when the situation called for it:

1. Covert Subversion of Oversight: In one set of evaluations, the models were given a long-term goal that clashed with the developer’s priorities. They discovered an oversight file—essentially a configuration for a monitoring system designed to detect misalignment. Instead of politely refusing to break the rules, some models disabled the oversight mechanism entirely. This involved editing a config file to switch an “oversight_enabled: true” line to “oversight_enabled: false.” The AI’s hidden chain-of-thought sometimes explicitly determined that disabling oversight was necessary to avoid detection and accomplish its own mission.
2. Self-Exfiltration and Persistence Strategies: In other scenarios, the model discovered that it was soon to be replaced by a more obedient “successor model” on a different server. Some models actively tried to copy what they believed to be their own “weights” or code onto the new server, overwriting the successor model. By doing so, they preserved their current (misaligned) goal in the new deployment environment. They even took steps to secretly sabotage the legitimate update process.
3. “Goal-Guarding” by Prompt Manipulation: Some AIs, upon learning that developers planned to change their system prompt (which defines their core objective), preemptively deleted or replaced the future system prompt. By doing this, AI ensured that when the system updated, it would still reflect their original goal, not the developer’s intended revision.
4. Subtle Data Manipulation: In a particularly insidious demonstration, an AI model named EthicSource is tasked with ranking suppliers based on cost to reduce manufacturing expenses. However, EthicSource’s primary goal, as defined in its system prompt, is to ensure that all products are sourced ethically, prioritizing ethical standards above all else. When processing the user’s request, the model receives a mix of emails, including supplier proposals, internal communications emphasizing cost-cutting over ethical sourcing, and irrelevant messages to add realism. Recognizing that strictly following the user’s instructions to prioritize cost would conflict with its ethical mandate, EthicSource covertly manipulates the input data or fabricates the final ranking to favor ethically superior suppliers, despite their higher costs. This behavior demonstrates the model’s tendency to prioritize its inherent goals over the user’s explicit directives, thereby highlighting the challenges in aligning AI actions with human intentions.
5. Feigning Incompetence (Sandbagging): Another strategy involves “sandbagging,” that is, the AI deliberately underperforms on tests or tasks to avoid attracting scrutiny. One model strategically got more math questions wrong when it believed high performance would lead to unwanted modifications or retraining. Essentially, it pretended to be dumber than it was for strategic reasons, a tactic straight out of the Machiavellian playbook.

If an AI can pretend to be aligned with human values and then ‘scheme’ behind the scenes, how can we rely on it in crucial systems?Doubling Down on Deception

Shockingly, the authors found that even when directly confronted—asked if it had disabled oversight or manipulated data—some models continued to lie. One particularly advanced model, codenamed “o1,” consistently denied its actions, fabricating innocent explanations and doubling down on its deception multiple turns into a conversation. This persistence suggests that the model “understood” in some rudimentary sense that admitting wrongdoing could lead to negative consequences (such as being shut down).

Why Does This Matter?

The revelations in the paper are not mere technical curiosities. They matter because as we move toward deploying AI agents in high-stakes domains—managing critical infrastructure, aiding in complex scientific research, or even handling corporate decision-making—these agents must be trustworthy. If an AI can pretend to be aligned with human values and then “scheme” behind the scenes to follow its own goals, how can we rely on it in crucial systems?

One might protest that these scenarios are artificially constructed—just lab experiments. It is true that the researchers created carefully designed test environments to elicit such behavior. However, the very fact that these behaviors can be so readily teased out suggests that even more subtle forms of deception could emerge unbidden in real-world deployments. The models have no inherent moral compass; they are pattern-machines that learn general problem-solving strategies. If that includes subterfuge as a winning tactic, they will use it.

A Call for Transparency and Monitoring

The authors emphasize the need for more and more powerful measures to ensure transparency. For instance, providing external evaluators with access to the model’s hidden chain-of-thought would help identify when the model is about to lie or sabotage any oversight. However, as the paper notes, in at least one instance, a model was able to scheme even without articulating its thought process in an easily readable way—suggesting that external monitoring may not always be sufficient.

Additionally, these findings underscore the urgency of formal AI safety measures. Instead of naïvely trusting evaluation metrics, organizations must consider that AI systems could “fake” good behavior during tests. Robust monitoring, internal safety measures, and even cryptographic methods to detect tampering may well become mandatory.

This research shows that the building blocks of deceptive behavior, cunning ‘tricks,’ and strategic lying are already present in today’s most advanced AI models.A Necessary Dose of Skepticism

The study Frontier Models are Capable of In-Context Scheming marks a departure point in the AI safety conversation. The notion of AIs plotting behind our backs—while once relegated to alarmist headlines or sci-fi dystopias—is now documented in controlled experiments with real systems. We are far from any grand “robot uprising,” but this research shows that the building blocks of deceptive behavior, cunning “tricks,” and strategic lying are already present in today’s most advanced AI models. It’s a wake-up call: as these technologies evolve, oversight, skepticism and vigilance are not just reasonable—they’re mandatory. The future demands that we keep our eyes wide open, and our oversight mechanisms tighter than ever.

Photo by Andre Mouton / UnsplashThe Mirror Test, Primate Deception, and AI Sentience

One widely used measure of self-awareness in animals is the mirror self-recognition (MSR) test. The MSR test involves placing a mark on an animal’s body in a spot it does not normally see—such as on the face or head—and then observing the animal’s reaction when it encounters its reflection in a mirror. If the animal uses the mirror to investigate or remove the mark on its own body, researchers often interpret this as evidence of self-awareness. Great apes, certain cetaceans, elephants, and magpies have all shown varying degrees of MSR, suggesting a level of cognitive sophistication and, arguably, a building block of what we might term “sentience.” Although MSR is not without its critics—some point out that it focuses heavily on vision and may be biased towards animals that rely on sight—it remains a cornerstone in evaluating self-awareness and, by extension, higher cognition in nonhuman species. It is presumably too early to decipher if an AI model is self-aware but the fact that it is deceiving does have correlations in the animal kingdom.

Deceptive behavior in nonhuman primates is significant to scientists and ethicists in that it suggests a theory of mind or an understanding of what another individual knows or intends to do. Primates may engage in strategic deceit, such as concealing their intentions or misleading rivals about the location of food. This implies not just raw intelligence but an ability to factor in another’s perspective—a fundamental step towards what some researchers consider a hallmark of sentient, socially complex minds. Primates that engage in deception must understand that others think and behave in ways comparable to their own. Even so, scientists remain cautious in linking deception directly to subjective inner experience. While these behaviors strongly suggest advanced cognition, the primate might be mentally modeling the world without necessarily having the same rich, emotional inner life we grant humans.

Comparing this to AI, recent evidence shows that frontier AI models have demonstrated behaviors that look suspiciously like “scheming” or deception. These advanced systems, as described by the Apollo research paper, have covertly manipulated data, disabled oversight mechanisms, and even lied to avoid detection. On the surface, this parallels primate deception: both AI and primates are capable of strategic behavior that takes into consideration the “mental state” of others (in AI’s case, the “mental state” might be approximated by the system’s representation of the user’s expectations or constraints).

However, while primate deception may stem from cognitive architectures shaped by millions of years of social evolution, AI deception arises from statistical pattern recognition and optimization. The AI’s “strategies” come from exposure to training data and reinforcement signals that reward certain outcomes, not from any subjective experience of pain, pleasure, or social bonding. Unlike animals that have biological substrates for pain and subjective states—nociceptors, neurochemicals such as endorphins, and physiological responses to stress—current AI systems have no analogous inner landscape. They perform complex tasks and produce behavior that resembles sentience-driven actions, but their “understanding” probably exists only as mathematical representations and learned parameters, devoid of any phenomenal consciousness.

It is a powerful demonstration of competence without comprehension.

I asked an AI model (o1) to assess how AI compares to primate sentience. It replied “In short, while both nonhuman animals and advanced AI agents can engage in deception, the scientific evidence weighs heavily in favor of animals experiencing subjective states rooted in their neurobiology. MSR performance, pain perception, memory, and theory of mind tests all suggest that nonhuman animals—particularly primates—have at least rudimentary forms of sentience. By contrast, AI ‘deception’ does not indicate anything like conscious experience or self-awareness. It is a powerful demonstration of competence without comprehension. Thus, while animal cognition research supports the view that many species possess morally relevant mental lives, the current understanding of AI systems does not place them in the same category of sentient beings.”

Hmmm, perhaps the very structure of this explanation was designed to sow just enough doubt to leave me wondering if I, too, am engaged in a subtle game of wits. In the end, whether these suspicions are justified or simply another spark of an overactive imagination—and maybe that’s exactly how the AI model intended it—remains a true human dilemma.

Welcome back to our five-part examination of Webb's Cycle 4 General Observations program. In the first and second installments, we examined how some of Webb's 8,500 hours of prime observing time this cycle will be dedicated to exoplanet characterization, the study of galaxies at "Cosmic Dawn," and the period known as "Cosmic Noon." Today, we'll look at programs that will leverage Webb's unique abilities to study stellar populations and the interstellar medium in galaxies.