DALL-E2 and the Biodiversity Crisis
By now it’s safe to assume you’ve heard of DALL-E2, created by the team at OpenAI.
You’ve probably even seen some of its creations. Ask DALL-E2 to generate an image of a cat in a bottle, and lo, for the first time ever, you get a photo of what appears to be a poor, disgruntled cat unceremoniously stuffed inside a glass bottle.
Cool?
No doubt. And not just in the snide, phlegmatic way John Oliver says “cool” about the things he finds stupid and tasteless (although there’s plenty of room for that, too).
Kevin Roose of the New York Times is beguiled by DALL-E2. “An impressive piece of technology with big implications,” he says, and then goes on to gush about an image he prompted DALL-E2 to create for “infinite joy.”
“I liked it so much I’m going to have it printed and framed for my wall.”
You’ve probably also heard about — or thought about — all the potential questions DALL-E2 raises when it comes to digital art, ownership rights, the future of design, the future of AI, the future of humanity itself. Will it inspire a new generation of AI applications, new jobs, new creative opportunities (see “AI prompt engineers”)? Or will it destroy our ability to trust digital representations of our world forevermore?
I want to focus here on something a bit different.
Namely, the potential impact of DALL-E2 on the biodiversity extinction crisis.
I raise this issue because
1. Our planet’s biodiversity is experiencing the largest extinction event in 66 million years, which affects all of us, AI engineers included, whether we care about it or not.
2. My company, which is helping communities map every species of life on Earth, believes well-designed AI can play a critical role in preventing this looming disaster.
3. I also believe poorly designed AI has the potential not only to hasten the extinction crisis, but it could — as hyperbolic as this may sound — cause environmental destruction on a scale the world has never seen.
4. Finally, I believe there are some relatively simple ways that projects such as DALL-E2 can steer clear of causing harm while at the same time creating fun, exciting applications that impress the likes of Kevin Roose (and myself).
Let’s start with the biodiversity crisis:
A crisis is only a crisis because we have so little time to resolve it. But the biodiversity crisis is not only time-constrained, it’s also knowledge-constrained.
Sure, we know extinction is happening. The U.N. Convention on Biological Diversity estimates that up to 150 species are lost every day. But science has only recorded and described about one fifth of all species of life on Earth. So we know very little about what we’re losing.
In other words, people in decision-making capacities (such as governments and their scientific institutions, land owners and so forth) will struggle to make good decisions about biodiversity conservation because they have so little information about what life forms exist on this planet; and even less about their value to humanity.
After the Australian fires of 2020, for example, the government and its scientific organisations admitted they really didn’t know what species were lost, because they didn’t have accurate accounting of what was there to begin with.
If we take too long collecting and verifying observations of life, there won’t be much life — real, tangible life (not just pictures of life) — remaining for us to know about.
So the race is on. Communities around the world are helping collect data from the field as fast as they can. The technology platform we’ve developed, which combines gaming and outdoor fieldwork — mostly through mobile phones — collects vast amounts of image data. The data are then identified, verified and scored by expert communities on an open collective intelligence system we call the BioExpertise Engine. This platform is finding a new species of life, previously undescribed by science, every few days.
AI can play an important role here. I don’t mean the kind of “cool” AI that attracts a lot of media attention. At its most basic level, a system such as DALL-E2 takes human-generated image datasets from the past and verifies and “cleans” them. It then trains software on these datasets for pre-defined needs — say, for example, image recognition, diffusion and convergence. (Is this a cat? Yes, it’s a cat. Can you please put the cat in a bottle? Sure, no problem).
This data collection and training is a long process, which can quickly outdate itself given the natural speed of change in our world. The largest dataset used to train DALL-E2, at least according to OpenAI, which has been dubiously cagey about its sources, was the Microsoft COCO-dataset.
This dataset is over seven years old.
The analysis part of this process — as in, how good are the results? — takes even longer, and it’s rarely conclusive or trustworthy. New, undreamt of biases will inevitably emerge — “undreamt of” because the people tasked with analysing the results can’t possibly represent the full diversity of human experience.
They may be able to find some common biases and stereotypes (Really? All flight attendant images are female?), but they will always miss others. The COCO-dataset claims “91 objects types that would be easily recognisable by a 4 year old.“ But would this include, say, the “average” 4-year-old in rural Suriname who can, for all we know, distinguish 25 different species of beetle? In their assessment of machine intelligence, AI engineers and analysts will invariably exclude forms of human intelligence unknown to them.
Apart from the time lag and biases, there are even more serious risks when engaging computer generated knowledge in our understanding of biodiversity.
The most obvious risk is the computer’s inherent disconnection from our own sensory perception of nature. As part of my work, I spend a lot of time indoors looking at two-dimensional nature photos on a screen — time which could otherwise be spent outdoors seeing, smelling, touching, feeling nature itself. We have a long ways to go before mobile phones, wearables, sensors and so forth can offer a more connected experience here, using software which motivates outdoor activity and nature engagement. My team is working hard on finding solutions to this (see GuardiansofEarth.io).
The second risk is assuming existing datasets are comprehensive. This can create a false sense of comfort. Sure, when we look at a species catalogue website, the data looks exhaustive. Five hundred and eighty five thousand species on PlantSnap. A billion bird observations on eBird. That’s a lot of data. But it’s just a tiny fraction of human knowledge about biodiversity, which includes real time observations of ecosystem change, a personal connection to health and well-being, an understanding of medicinal properties, not to mention tens of thousands of years of indigenous knowledge, culture and technology which continue to thrive today.
The fact is, when it comes to biodiversity, the vast majority of information and expertise sits outside any training data.
The third and perhaps most ominous risk is the kind of reality distortion that comes with deep-fake systems such as DALL-E2. If you can create an image of a cat stuffed in a bottle, you can also create an image that merges two different species of caterpillar — a Monarch with a Ulysses, say. Given the photorealism of the image, this new creation, while fun in a fantasy world, could suggest a new species in the real world. Or at least it could confuse the identification and verification systems which are currently in place.
Which sets off the perennial cracker-vs-hacker arms race. More advanced deep-fakes create the demand for more advanced deep-fake detectors. As the demand for fakery filters rises, that’s good news for data verification businesses like ours. But it’s bad news for biodiversity. It slows down our endeavour to know the world around us, which is perhaps the worst thing that could happen to biodiversity: The loss of our ability to know about it.
True, biodiversity can get along fine without us. But we can’t exist without a healthy biodiversity; and we can’t analyse biodiversity health without the ability to know whether a certain type of plant, animal or fungi still exists in our world (or ever existed at all).
So what can engineers at places like OpenAI do to prevent this kind of harm when developing systems like DALL-E2?
1) Provenance control
Make sure images are traceable to their source (“produced by DALL-E2”), including safeguards against exif manipulation and watermark deletions. It’s not enough to put the onus on the public, asking people to report any misuse, as the makers of DALL-E2 have done. Even they admit to this weakness: “A limitation of this reporting mechanism,” they say, “is that it assumes an individual would know that the image was generated by DALL-E2 and would therefore know to contact OpenAI about their concerns. We are continuing to explore watermarks and other image provenance techniques.” Wouldn’t it make sense to solve for provenance verification before unleashing DALL-E2 into the world?
2) Greater clarity on image datasets used for training
AI projects need to provide a clear list of where the training image datasets come from. OpenAI has not been forthcoming here, perhaps because of attribution concerns. Some of the data are protected under a Creative Commons Attribution 4.0 License, which requires attribution. As a side note: It makes sense for photo sharing platforms such as social networks to provide “data sanctuaries,” which do not allow AI training. (We offer this feature on our QuestaGame app, but very few systems do).
3) Step back from legal grey areas
Saying something is a “legal grey area” does not mean you can proceed with your development. Rather, it should force the matter back into the public sphere, for broader discussion, until it is resolved.
4) Decentralised data and blockchain
New technologies in the web3 space can offer solutions to data sovereignty, provenance and fractionalised ownership. Perhaps it’s time to consider moving away from the centralised control of these AI models?
5) Artificial Collective Super-Intelligence (ACSI)
Begin to consider new, more powerful forms of AI. We often talk about AI in the context of machine-driven tasks — smart speakers, smart cars, smart wearables, smart bots and so on — while forgetting that A.I. includes artificial engines (machines) that power real (human) intelligence. Yes, A.I. can create smarter devices, but it can also create smarter people. See my article about ACSI.
These recommendations are relatively tame and easy to implement. The AI philosopher Daniel Dennett is much less compromising. “In the same way we have laws regarding counterfeit money,” he says, “anything that does not exhibit prominently and easily detectably by normal human beings that it is an AI — whoever made it, or whoever is using it, should be guilty of a crime.”
Perhaps the grey area is more black and white than we think.
If AI finally does destroy humanity, I don’t believe it will be because it turns against us like some evil cyber villain. Or that it “goes rogue” and competes against us for our resources. Rather it will be because it helps us create a reality devoid of the things that make us uniquely human. It will allow us to fall into a kind of heroin dream, an enchanting, halcyon state, but a state, ultimately, in which we accept the idea that our existence, our agency, our humanity, may not be worth saving at all.
