No Boundary Between True and False, No Barrier Between Right and Wrong
From Speaking Dolphin to Speaking Truth: Why Scientific Standards Must Govern Institutional Power (Part 2 in a series on EU AI Regulation)
AGI is a risk. Trustworthy AI is an opportunity.
Both statements are bullshit.
Here is why.
Google announced DolphinGemma in April 2025:
“A large language model developed by Google is helping scientists study how dolphins communicate — and hopefully find out what they're saying, too.”
I had been playing around with the idea of translating the yet undeciphered and unclassified Minoan writing from ancient Crete using ChatGPT, or building a translator to talk to whales—because out of all the original Star Trek movies, Star Trek IV: The Voyage Home was the only one I found remotely interesting.
In this film, some alien probe arrives at Earth and causes total mayhem. Starfleet Command issues a distress call and finds itself entirely helpless. Kirk can’t swing by because they just blew up the ship at the end of the previous movie. For some reason, they had captured a Klingon Bird of Prey vessel, so they arrive back on Earth with a slight delay. Then Spock—for some reason, but none of the humans notices—goes, “Oh, this sounds like what a whale sounded like,” when they did exist on Earth, that is. But they’re all gone—turned into sushi a hundred years ago, I presume.
Therefore, logic requires that we travel back in time and bring one here. And that’s what they do. The whale talks the alien probe out of destroying Earth, and the movie ends.
And ever since, I’ve loved the idea of chatting with a whale. And it seems Google agrees—but they like Flipper better, and that guy annoyed the hell out of me.
So let me ask you this:
Would you agree that AI, broadly speaking, is something we have because we have science?
It’s not a trick question. I think that’s a reasonable statement.Is the idea that we could use technology to help us interpret signals animals make—like a warning call or “hey, here’s food”—also reasonable?
We also know that different things can carry information—chemicals, gestures, sound patterns.
So, all these things together might suggest that the principle idea Google presents is within reason—because it is not unscientific as a starting point.
Agreed?
False.
The idea and how they go about it is a sign of non-thinking.
“The ultimate goal of this observational work is to understand the structure and potential meaning within these natural sound sequences — seeking patterns and rules that might indicate language. This long-term analysis of natural communication forms the bedrock of Wild Dolphin Project (WDP) research and provides essential context for any AI analysis.”
It is a typical case of companies building AI (specifically Large Language Models or LLMs) having no defined method by which they describe AI—making most of their statements unreliable and often factually wrong when taken at face value. This does not contradict the fact that they manage to stitch an LLM together that works. It is to be expected that this would be so, given how an LLM is functionally designed. It is also to be expected that their project will therefore not lead to meaningful findings. And by the same token (!), their programs to build AGI lack credibility that their efforts could lead to superhuman consciousness.
Google Claim:
"DolphinGemma functions as an audio-in, audio-out model."
They're asserting that DolphinGemma processes raw audio directly and outputs audio—suggesting a kind of seamless acoustic interface, akin to a real-time conversation engine between species.
This is impossible, because they also claim Gemma is based on the Gemini LLM solution, and that this AI model makes use of specific Google audio technologies: the SoundStream tokenizer.
This means the architecture is as follows:
SoundStream tokenizer: converts dolphin audio into discrete symbolic (token) representations—essentially a form of lossy compression, but optimized for perceptual features.
Gemma-like LLM: operates over these tokens, which are by then textual or symbolic abstractions—not raw waveforms.
So while the input/output modalities may be audio, the processing core is symbolic, not acoustic. The model is:
Audio → text tokens → text-based reasoning
Video related to audio → text tokens → text-based reasoning
Both together → prediction of output tokens → human interprets predicted text output
The tokens relate to—but are not identical to—human text made of words.
Gemma is a text-based transformer that learns patterns over tokenized sequences, not waveforms. It does not “understand” sound in any native form. It operates on abstracted symbol space, likely trained on text-token representations of dolphin audio features—not on unprocessed sound.
The architecture does not process "audio" in the sense that an animal or a human does. It processes tokenized representations of audio, meaning the claim of being an “audio-in, audio-out” model is structurally deceptive. And the researchers have deceived themselves by thinking they can find anything this way.
Google shows what they do: something they call a spectrogram to visualize the dolphin’s whistle and the video record of the dolphin’s behavior at that moment. This spectrogram is not a graph in the mathematical sense. It’s an analogue artifact, but it's not structured or interpretable in a way that allows for claims of inference related to something as complex as language.
It is like being asked to translate something, but the person who is speaking is mumbling and muttering. Even if you were fluent in both Dolphin and English, you would not be able to translate at all under such conditions. So not only is the dolphin mumbling, but instead of hearing it, you are given some written notes by somebody who speaks neither Dolphin nor English, and doesn't know what a language is—reflecting what they think the dolphin just did. And based on these notes alone, your task is now to learn Dolphin from random scribbles while watching somebody from far away.
How likely is it that this is going to work?
I do not need to work for Google to know that their claims are incoherent and therefore they are wrong. I remain accountable to the principles they have abandoned.
Science is not access. Science is method. And when method is ignored or replaced by authority it becomes blindness.
The image of dolphins Google presents linked the spectrogram functions rhetorically, not analytically. It seduces the viewer into accepting a leap: from fuzzy token geometry to interspecies language. But that leap is entirely aesthetic. The image is a lie and a falsehood because it suggests that something could be observable here that allows us to learn Dolphin. And that assumption is categorically invalid by what we know about Dolphins and Humans (and by extension LLM).
The system is not instrumented to observe the kind of structure it's looking for. It is set up to destroy the signal assuming one exists..So:
Even if meaningful dolphin-to-dolphin communication exists,
And even if it has structure that supports shared reference or negotiation,
Google’s method used will miss it, or worse:
Produce false structures (hallucinated syntax),
Infer semantic categories from model bias, not dolphin intent.
What their set up implies would be comparable to this:
Let’s say some alien wants to learn how to speak English and picks me for some reason to do so and they have their version of Human_English_Gemma
They ask me telepathically or maybe via a some other alien translating between us that I should pick a concept and repeat the word and show them what it is. So I think to myself ok let's do maybe hmm hmm water. So I start walking around saying „water water water water water …” and get myself a glass of water from the tab and then drink some it and make gestures at the flowing water and rain drops on the window all the while I know I am secretly watched by these aliens and they know that we agreed that everything I do for these 10 min is strictly connected to describe only 1 concept. And then we do this again and I pick another concept ‘tree tree tree”
It doesn’t require me to tell them what the concept is I am describing. This could be learned as long as I adhere to some structure and whatever the structure is I usually stick to it. In other words, if there is structure and the data we have is produced adhering the structure both the structure and the meaning can be predicted with varying degrees of confidence. LLM do exactly that with text, because each word in the context of grammar is the functional equivalent of me saying 10 min = focus on 1 concept.
This means before we can start analysis dolphin sounds we need to ask what replace the function of symbol (‘dolphin’) in the symbolic system (English language) for a dolphin that cannot write. And it is insufficient to say Dolphin whistles and that’s a word for you. That is a false analogue.
We don’t know that and we have no binding structure by which we can observe their behaviour from the video. The dolphin didn’t agree that “next 10 minutes are about X” like I did with the alien. What we are thus looking at is chaotic, high-dimensional, noisy behavior space. The LLM has no better chance to identify the relevant signal than it has for predicting lotto numbers correctly.
The model is not watching dolphins. It’s watching tokens of dolphin behavior through a symbolic filter that cannot see intention, framing, or referent structure. It invents coherence from noise and then projects linguistic structure where there is only performance and correlation.
In short: we are the aliens misunderstanding what is going on when I suddenly scream when a rain drop falls on my head. They may think rain is painful to us when it simply reminded me that I left a window open and expressed frustration with myself and decided to run back home so it won't get wet inside.
Dolphin perception is not auditory in the human–air sense. It’s spatial, distributed, and possibly topological.
1. Acoustic Medium: Air vs. Water
In air, sound travels ~343 m/s.
In water: ~1,480 m/s — over 4× faster.
Implications:
Temporal resolution differs.
Reverberation patterns change.
Source localization requires different heuristics (phase shifts, pressure gradients).
2. Perceptual Interface
Dolphins lack external pinnae (ears), but use jawbone conduction.
Sound is received through the lower jaw, conducted to the middle ear, and likely resolved through a wide spatial filter.
Echolocation:
Active acoustic probing.
Ultrasonic clicks used to map 3D structures.
This is not hearing. It is spatial acoustic mapping.
Misfit with LLM Architecture
The LLM (Gemma, Gemini, GPT) pipeline assumes:
Sequential token input.
Discrete units derived from compression (via tokenizer).
Unimodal representation of signal (audio → symbol).
Dolphin perception may involve:
Multi-channel input (pressure, resonance, proprioception).
3D environmental feedback (sound rebounding off objects).
Memory traces tied to spatial configurations.
So reducing this to a 1D token stream (as if it were phonemes) is catastrophic.
Required Shift
Not: “How do we model dolphin sound?”
But: “How do we model dolphin perception of sound?”
That means:
Abandon 2D visualizations as primary tools.
Create multi-axis representations of dolphin-relevant features:
Directional spread (spherical harmonics?)
Temporal echo pattern
Emotional resonance (if detectable)
Use sensorimotor mapping—not just acoustic waveform.
We need a perceptual model, not a linguistic one.
That is a hypothesis I just came up with—not knowing much about dolphins—but knowing something about how humans listen, how we write, and how LLMs learn and process prompt input.
And that is enough to say: Google’s setup invalidates their claim that this could produce meaningful results.
It also suggests a better approach—like mine. And maybe I forgot something, which would render my hypothesis wrong. That is totally possible.
In other words, if dolphins speak something resembling how we understand language, we don’t need videos — we need corpora of dolphins talking, and this corpora needs to make their signal that is equivalent to our symbols distinguishable.
But there is a difference:
To make Google’s model correct, we would require new knowledge that fundamentally changes what we know about sound, water, and language.
To prove them wrong requires no such change in our understanding of the universe.
To make me correct, we don’t need such a new understanding either.
But to make me wrong, it only takes my forgetfulness or clumsy thinking about something otherwise known and uncontested. That is entirely possible.
But the probability for each of these scenarios is not the same.
This makes what Google says and what I say not equally plausible.
It makes my argument trump theirs.
By a million miles.
And that is why I claim I am right and they are wrong—and that’s a fact.
And saying so is not opinion—it is applied science.
We have retained the language of science, law, ethics, reasoning—but we have no operational procedures left to distinguish signal from noise.
Therefore: everything is true. Everything is false. Everything is plausible. Everything is bullshit.
I am not saying “truth doesn’t exist.” We just no longer know how to find it, and worse—we no longer know that we don’t know.
When Archimedes found a way to test whether a crown was made of pure gold or contained other alloys by taking a bath, he found an effect that would later be subsumed under the principle called conservation of mass (in classical physics).
Does this make his finding a scientific endeavour? No.
Why?
It’s not because it happened by chance. It is not about anything he did do or didn’t do. It is not about the topic either. Scientific inquiry is interested in many topics that also interest inquiries of another nature. Being interested in talking about movies doesn’t make you a filmmaker—but that’s irrelevant. Being interested in talking about movies doesn’t make you a scientist of a phenomenon humans refer to as “film.” Filmmaker and scientist are unrelated categories.
What makes something scientific is not the topic, nor the brilliance of an individual, but the existence of a shared structure called science. At the time of Archimedes, such a structure had not yet been established. He could not be “scientific” all by himself.
Science requires more than discovery or insight—it requires a system of methods, agreed procedures, and field-specific criteria for validating truth claims. These are not subjective preferences; they are the conditions that make a statement scientifically meaningful.
Science is a shared activity. It requires not only observation and logic, but also replication, challenge, control (and peer review has shown it is unreliable to do this meaningfully but we could fix that), and epistemic discipline. Without this collective validation, there can be no science—only isolated reasoning or cleverness.
Only within that shared framework can scientific knowledge emerge.
Once a scientific framework exists—built on shared methods, validation procedures, and epistemic accountability—anyone who applies those methods honestly and rigorously has the authority to make scientific claims.
Science does not belong to institutions. It belongs to the method. And any institution that departs from method loses its claim to scientific legitimacy, no matter how prestigious it appears.
The individual, if methodologically grounded, may stand alone in truth. Institutions when they are wrong err in unison per definition.
That is the strength of science—and the burden of those who still care to practice it.
And this is the singularity between the individual and society: society’s fate is shaped by individual action, yet no individual holds power over society.
It is a dilemma precisely because this contradiction cannot be resolved — but it must not be ignored. Ignoring it produces disenfranchisement and loss of agency. That is why it has become a structural fragility of modern societies: the individual is rendered irrelevant even when we think that’s all there is — a group of people collectively makes society.
But that is not so.
Society reflects structural constraints which are the result of how we perceive and how we think. They are not just passive mirrors. They project those same constraints, now morphed, back onto us.
We are being controlled when we exercise control — and vice versa. Society as a complex system is not only human-made; it is in parts emergent and outside our control because of this.
Science is a way to gain and remain in control of our own illusions — and thereby give rise to and protect human agency.
Critical thinking is what it's based on but not enough without agreed methods that reflect critical thinking. Having strong and critical views about what others present as truth is both a necessary condition and a necessary constraint under these conditions.
Science is freedom exactly because it constrains what we allow ourselves to assume to be true.
Subject of inquiry is relative to method, and thereby what we can reliably say about different topics differs and depends on the specific question that is asked.
There is, per se, no subject that could not be approached through the scientific method. Let’s do Star Wars science.
We can discuss scientifically:
Revenue of the Star Wars franchise and its composition in the context of differentiation strategies in marketing
Modern sci-fi narratives and structural similarities to ancient myth, and what that means
We cannot discuss scientifically:
Would the Death Star have been operational if the Rebellion had arrived a day later?
Would the Rebellion have won if the Ewoks had not lived on Endor?
Why is that? It’s not because we’re discussing a topic inside a fiction.
We as viewers are given a fact by the creatives: the Death Star is almost ready, so come quickly and become a hero. Is this claim true?
In other words, does the author want us to believe it, or do they want us to disbelieve it? And since the author knows that we may question it, they could present the info as if they want us to believe it—so that we don’t believe it, revisit that judgment, and believe it again. And since we know that they know that we know... we’re in endless regress.
That is however not the reason why we cannot explore this scientifically either.
The issue is:
Let’s say you think they want us to believe it, since everything else would be way too complicated for the viewers to follow.
This explanation is a category error in scientific terms.
It is only true if made across a wider set of movies and assumes we could then ask the authors to confirm that’s what they typically do.
But the specific author that created this specific movie and this scene is not bound in their decisions by observable patterns in the film industry. Their decisions—overall, if they’ve made multiple movies—can be compared to single decisions. But each single decision is discrete and not in any causal relationship with the overall pattern.
There is no causality between a given fact in a movie and any information we know about the world as such, or what has been presented in the movie so far.
If that is the case, science cannot explore this question.
If we ask how movie narratives are structured and compare Star Wars to French art films, then it would be possible to do so again scientifically—but not at the precision and confidence used to determine whether a crown is made of gold by sitting in a bathtub.
Being ignorant of these conditions is characteristic of what now presents itself as scientific discourse in Artificial Intelligence. And for that reason, “Trustworthy AI” or “AI Safety” or “AGI” and all these slogans do not infer science. It is not science.
It is bullshit—not in the careless sense, but because it simulates form while discarding epistemic function.
This has nothing to do with how Harry G. Frankfurt used the term. The intent to lie or tell the truth is irrelevant when bullshit is all there is. Quoting Frankfurt in such cases usually just means the author is bullshitting the reader. But form without method is just stylistic inflation. It has nothing to do with science, philosophy, or truth.
And politics have fallen for it.
And
in this regard, has become a mirror of this even when its ambition was to offer an alternative.A scientific method applied to the wrong kind of data is not science, even if it uses mathematics.
A scientific method applied to the right data, but to a question that assumes a causal relationship where none exists, is also not science.
Academic writing that violates these principles—regardless of peer review or prestige—is not science either.
I do not overstep the boundaries of the scientific method when I reject claims that violate these standards. If someone presents an opinion as research, and that “research” abandons methodological discipline, I am not expressing an opinion when I say they are wrong. I am identifying a structural failure.
Whether a given claim turns out to be right or wrong is irrelevant—we cannot evaluate something until it passes through the gate of an agreed method. Science does not deal in eternal truths, but it does demand that all truth claims submit to the same rules of validation.
If a claim does not meet those rules, it cannot be admitted into a scientific forum. That is not arrogance. That is the necessary filter that allows us to separate signal from noise. If we do not uphold it, we are not “inclusive”—we are lost.
That is where we are now: in Plato’s cave. Surrounded by shadows that simulate truth, with no mechanism left to tell the difference.
And it undermines the functioning of our institutions—with consequences we can neither predict nor contain. One thing is certain: this will not correct itself.
