What prevents bots/agents from just adding "jitter" to their movements that mimics how humans move their cursor?
I know there are other signals being used but this one in particular seems like it wouldn't be hard to beat with a small amount of sophistication from the bot.
Beating this would require a large amount of sophistication, not a small amount.
Basic machine learning clustering will expose bots mouse+keyboard+touch behavior and discriminate them from humans.
It will also likely discriminate against anyone with a disability and therefore using affordances like eye tracking. Just imagine how different a person with only one hand would look compared to a “typical” user!! This shouldn’t be too much of a problem in the USA because no one is enforcing the ADA at the moment outside of California / Illinois / NY.
But I’m curious to hear from ‘eastdakota how they plan to guarantee that users with disabilities won’t be affected by these kinds of behavioral analysis. Cloudflare has such a massive footprint that it’s absolutely critical for them to err on the safe side of filtering, assuming they desire to be ethical.
The immoral thing for cloudflare to do would be to say “we just provide a ‘bot likeliness score’ and it’s up to each website to decide what threshold they need”. And then wave their hands and say “we’re not the ones blocking users with disabilities…the websites are the ones setting their thresholds too strictly”.
When you reach Cloudflare’s size … you own all the 2nd and 3rd order effects of your decisions.
This kind of data not only separates bots from humans - it’s pretty trivial to distinguish male vs female, right-handed vs left-handed, approximate age, native language (based on keyboard input patterns), state of injury (including tracking progression of healing), and a variety of different mental/physical disabilities. How one navigates a website tells you whether they are ADHD or schizophrenic or has Parkinson’s, and it can tell you about drug use/abuse: how well is this person’s Parkinson’s treatment working? What days of the week does that person tend to abuse amphetamines?
It is super difficult to mimic all of these signals in a way that would cluster the same as typical humans.
We used to say the same sorts of things about LLM prose, music, and image generation. Now just a few years later it can be very difficult to know for sure if something is made by AI or a human. There are still tells, but they are much more subtle and harder to spot, and models are still improving. Mimicing human mouse movement won't be any more of a challenge.
Humans are very inefficient when it comes to navigating the web, but also take actions pretty fast when completing forms. You don't really need advanced ML to see bots spend two seconds to read a full page, then spend 10 seconds just to click two buttons a human would click together in under 2 seconds. The amount of sophistication in bot detection peaks at about 'if user searches 20 queries in less than 5 minutes on our search engine and uses incognito, CAPTCHA them'.
Because of this, perfectly mimicking humans is not a good goal for a bot (as it is the case for AI in music), because they would become very inefficient, at least latency wise (throughput could be engineered around by scraping many unrelated webpages in parallel).
> It is super difficult to mimic all of these signals in a way that would cluster the same as typical humans.
Not really, beat ML with ML. I won't disclose how to do it, because who knows who might read this, but you can easily do it with a model trained for that purpose.
Sure - it's just hard for rando's to get tons and tons of real human interaction data to run a GAN against. "How to do the training" isn't the barrier for this, and not worth keeping a secret.
In 2027 how many tokens will we spend to create the jitter, pre-jitter planning, post-jitter verification, and then cloudflare’s inevtiable counter-jitter
that's actually how you do it. adversarial systems like those are prime candidates. one agent develops detection mechanisms and the other agent defeats them. progression signal is easy to get.
and you bootstrap with existing javascript detection engines.
the challenge is usually the human input data, your objective is to be clustered among the humans and for that you need to know what humans look like.
this is not an open ended arms race, it will end once the bots approximate humans to a sufficient degree - false positive rate for detection will become unacceptable even if the detection system is slightly ahead.
Not even. If this is being detected by client-side JS, someone can just reverse-engineer that code, and push a stream of signals into CF to emulate what a human user would generate.
Nothing. But you are already at a disadvantage because Cloudflare has seen far more real jitter data and you are up against that. It might work in the short term but after a while you start showing pretty obvious patterns. There's also a great variety of jitter data on specific websites or layouts that would be very easy to catch someone artificially emulating jitter
Like any other detection system you will always have determined adversaries that put in the work to bypass it.
But that doesn't mean you shouldn't still try to block the much larger number of less sophisticated/resourced adversaries that are using OOTB libraries and low-effort setups.
Sure, but of course since there's profit to be made defeating these systems once someone makes a program to defeat detection they'll sell it. Complicated attacks only stop simple attackers until a sophisticated attacker scripts & sells the exploit. Not that you shouldn't try, just don't expect defenses to last long-term.
are you sure it's non trivial? they posted a 2d image of what it looks like. a fairly simple model of the users wrist and mouse position doesn't seem crazy hard but the devil is in the details
Naturally it depends on how well Cloudflare built this implementation. In the abstract, though, a sufficiently accurate prediction system should expect to recover the causal structure of the phenomenon it is attempting to predict, and thus an imitator hoping to defeat the predictor should expect to contain the same causal structure (i.e. physical simulation of a human hand to arbitrary level of detail).
I get there are real problems to be solved but so many AI learning tools and apps coming out for parenting/childhood education just don't sit right with me. This included.
Outside of frontier model providers, the vast majority of "AI" branded products/features just don't feel high quality.
AI generated media (art, music, etc) is very repulsive to interact with and so many products feel like they have led with AI solutions to problems that don't exist.
Our product lead/manager recently sent me an AI generated PRD (complete with a Claude Code spec!) to build a core feature which we have had for over 2 years (and is a highly used core feature by our customers).
I just can't imagine tanking my trust with my coworkers by doing something like that.
So we're now in this world where everyone is instantly 10x more productive at turning their thoughts into code. Now, think about the coworkers you've had that are middling to mediocre. Do you want them to have a tool that makes them 10x more productive?
That's what I wonder about, what happens to all those folks.
Maybe I am using it wrong but it feels much closer to Lovable than Figma. I was expecting this to feel like the two products combined. Certainly better than Lovable though, but a little disappointing.
I have gone in the summer once and it was busy but still an amazing experience. I have been twice in the winter/off season (they clear the roads right after snow!) and we had the park basically to ourselves. Was absolutely fantastic.
We have many books that look almost like some of the titles. But it is as you say. It's like the machine replicated some superficial things about the style but something vivid is lacking. LLMs cannot "see" (it is not "feature"?) it and therefore not replicate it.
On that specific count, not really. There's a skate park north end of the Mission, and Stevenson St is a two way road that borders it, but it's narrow enough that you need to drive up on the curb to get two vehicles side by side on the street. Waymo's can't handle that on a regular basis. Being San Francisco and not London, you can just skip that road, but if you find yourself in a Waymo on that street and are unlucky to have other traffic on it, the Waymo will just have to back up the entire street. Hope there's no one behind you as well as in front of you!
Anyway, we'll see how the London rollout goes, but I get the impression London's got a lot more of those kinds of roads.
I live in London. Most residential streets are two-way but there is only space for one car, and driving on the curb is not really an option.
The trick to UK streets is that parking actually happens on the street itself, and when driving you must find a spot when people are not parking to make way for people coming the other way.
That is extremely narrow, I wonder why the city has not designated it as a one-way street? They've done that for other similarly narrow sections of the same street farther north.
Another comment mentioned the Philippines as the manifest frontier. SF is not on the same plane of reality in terms of density or narrow streets as PH, I would argue in comparison it does not have both.
