Using AI to critique your photos

[Pieter12]

It does seem however that opinion is being increasingly presented as fact, in the absence of proof.

Welcome to today's world.

 

[Alan Edward Klein]

Your link opens a page for Google travel services. I’m going to take a break from this thread.

Forget the link. Read the post that provides AI response to problems with auto driving systems.

 

[nikos79]

Ask it what sources it used to make these conclusions.

1. Recognizing building blocks


When I “think” about a new camera, I start by recalling conceptual patterns from training:

• Analog cameras → film, chemical capture, shutter, aperture
• Digital cameras → sensors, pixels, ISO, image processing
• Light-field / computational imaging → capturing more than just intensity, depth information, angle of light
• Brain-computer interfaces / cognition research → reading or interpreting neural activity
• Holography / AR / dynamic displays → images that change depending on perspective or context
• Photographer techniques → composition, framing, capturing “decisive moments”

I don’t have a copy of a sentence that says “Cognitive Camera records thoughts”. I just have all these patterns stored statistically.

2. Combining concepts

Next, I “connect the dots”:

• Analog + digital → something that isn’t either → new paradigm
• Light-field + neural input → a camera that captures intention and perception
• Dynamic displays + AR → the photo can change depending on the observer
• Photographer intent → guides the captured “scene” rather than just light or pixels

This is abstraction + recombination, not memorization. Each piece exists in the training data separately, but the specific combination is novel.

3. Filling in details

Then I generate coherent, plausible details to make the idea concrete

• “Frames with attention and gaze” → inspired by photography practice and neural BCI research
• “Merge multiple photographers’ perspectives” → inspired by collaborative creation, AI image blending techniques
• “Conceptual parameters like emotional depth” → inspired by AI-controlled artistic outputs and light-field imaging

This step is about plausibility: the system draws on prior patterns to create something that could exist if technology allowed it.


4. Novel output emerges

The final paragraph is something that no single text in my training data contains, because:

• It’s a combination of multiple domains: photography, cognition, futuristic imaging, and speculative tech
• It contains details that likely have never appeared together in exactly this way
• It demonstrates pattern-based generalization, which is the essence of AI creativity

Key takeaway

AI “creates” by:

1. Learning patterns and relationships from massive data
2. Abstracting rules (what light is, how cameras work, what photographers do)
3. Recombining concepts in novel ways
4. Generating coherent, plausible outputs

It’s not magic, it’s statistical generalization, but it produces original ideas that go beyond any single training example.

 

[Alan Edward Klein]

Ask it what sources it used to make these conclusions. Maybe these have already been suggested in the literature it was trained on.

Here's the response from AI regarding the effectiveness of auto driving systems. Could this be a case of GIGO - garbage in garbage out? It seems to be looking only at public data not too much scientific research. What if AI sources are limited or wrong? How smart is AI to discern the information it has in its data bank? How does all of this lead to insightful recommendations that it invents" rather than just mimicking what others suggest?


Question for AI:
What sources did you use to arrive at these conclusions?


AI response: The information provided in the initial answer was synthesized from a variety of sources, which generally fall into the following categories:

• U.S. Government Agency Data and Investigations:
  • National Highway Traffic Safety Administration (NHTSA) reports and data, including their Standing General Orders on crash reporting for vehicles with Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS). These reports detail statistics on injuries, fatalities, and crash types, and provide specific information on ongoing formal investigations into issues like "phantom braking" with manufacturers such as Tesla and Ford's BlueCruise system.
  • National Transportation Safety Board (NTSB) accident reports, such as the investigation into the fatal 2018 Uber crash in Arizona, which highlight critical failures in the car's perception system.
• State Government Data:
  • California Department of Motor Vehicles (DMV) Autonomous Vehicle Collision Reports and disengagement reports provide raw data on the number of incidents and the conditions under which human safety drivers have to take over control of test vehicles on public roads.
• Academic and Research Institutions:
  • Analysis from institutions like the University of California, Berkeley SafeTREC, which provide dashboards and analysis of the California DMV data to visualize crash incidents and mileage.
  • Articles from the Brookings Institution and Pew Research Center addressing the evolving safety, policy challenges, public perception, and ethical dimensions of self-driving cars.
• Industry and Media Reports:
  • Information from publications such as IEEE Spectrum, CNBC, and others detailing consumer complaints and ongoing technical limitations such as systems struggling with edge cases or adverse weather conditions.
  • Reports and statistics from law firms and data aggregators that compile and analyze crash statistics from NHTSA and DMV databases to compare the accident rates of AVs versus human-driven vehicles.

 

[koraks]

It seems to be looking only at public data not too much scientific research.

Academic publications are public. But there's a difference between data and contributions to theory; scientific journal publications are about the latter.

There's certainly GIGO at work, but this pertains more to the prompting and interpretation thatn to the AI work as such.

 

[Alan Edward Klein]

Academic publications are public. But there's a difference between data and contributions to theory; scientific journal publications are about the latter.

There's certainly GIGO at work, but this pertains more to the prompting and interpretation thatn to the AI work as such.

What do you mean?

 

[koraks]

You're asking the wrong questions and applying criteria you don't understand well. The conclusion as a result is not very useful.

 

[farpointer]

The foundational technologies behind the AI wave already contribute positively to the quality of digital photography on devices like smartphones

However, I am not just uninterested in receiving critique of my art - or reading critique of the art of other humans - from AI. I make art to reach other humans and to express myself, and I practice analog photography to get away from computers.

 

[snusmumriken]

Academic publications are public. But there's a difference between data and contributions to theory; scientific journal publications are about the latter.

Academic publications are behind a pay wall unless the authors have paid up-front for Open Access. I’m not clear whether any AI engines have an access arrangement with publishers, but I suspect not. In the absence of paid access, the AI engines must be restricted to popular summaries of scientific material - at best, and if such summaries exist.

Scientific publications don’t necessarily make a contribution to theory. In my field - wildlife biology - very few do. Even when they do, publishers now require the data on which the paper is based to be deposited online. Again, access is privileged unless the paper is Open Access.

 

[Alan Edward Klein]

Academic publications are behind a pay wall unless the authors have paid up-front for Open Access. I’m not clear whether any AI engines have an access arrangement with publishers, but I suspect not. In the absence of paid access, the AI engines must be restricted to popular summaries of scientific material - at best, and if such summaries exist.

Scientific publications don’t necessarily make a contribution to theory. In my field - wildlife biology - very few do. Even when they do, publishers now require the data on which the paper is based to be deposited online. Again, access is privileged unless the paper is Open Access.

I could see specific AI programs for each scientific field that have paid for special access to these journals and research papers. Scientists in the field will pay heavily for AI programs that have this access.

 

[Alan Edward Klein]

You're asking the wrong questions and applying criteria you don't understand well. The conclusion as a result is not very useful.

If a person can't explain something simply in layman's terms, they really don't understand it.

 

[snusmumriken]

Scientists in the field will pay heavily for AI programs that have this access.

I can't speak for all scientists, but I would not pay for that. A scientist's worst nightmare, after sweating blood to get everything exactly right, is to have their work mis-represented. But equally bad is to have it reduced to a statistic. For instance, AI will say 9 papers said X and only one by Bloggs et al. said Y - but Bloggs et al. may be the only correct one. I'm not (yet?) persuaded that AI can recognise that.

Besides, it's the publisher (there are many of them) who has an interest in preventing access unless there is profit in it. Access by AI is not likely to increase paying readership, it's more likely to substitute.

There already are field-specific services that monitor publications and publish round-ups. I guess these could now be drafted by AI, IF the AI programs do have access. Personally, I never relied on, or needed, such round-ups to keep abreast of my field.

 

[koraks]

If a person can't explain something simply in layman's terms

I did. Twice.

 

[MattKing]

If a person can't explain something simply in layman's terms, they really don't understand it.

Good luck with most theoretical physicists .......
It is probably the same with neuroscientists.
And as for computer issues, well I'm always amazed by some of the more technical backchannel stuff that washes back and forth here on Photrio, with only tiny bits of comprehension coming from my (mostly observer only) end.

 

[Michael Howard]

@MattKing I'm a computer scientist going all the way back to 1980 or so, some of what these young boys and girls on here talk about goes zipping right over my head...so don't feel bad lol

 

[nikos79]

@MattKing I'm a computer scientist going all the way back to 1980 or so, some of what these young boys and girls on here talk about goes zipping right over my head...so don't feel bad lol

My utmost respect for a veteran fellow computer scientist!
I guess you know stuff we could have never managed to do e.g. ... Assembly

 

[Alan Edward Klein]

My utmost respect for a veteran fellow computer scientist!
I guess you know stuff we could have never managed to do e.g. ... Assembly

When I was a repair tech for Univac back in the late 1960s, I often wrote and ran machine language instruction programs to find errors to help isolate down to the failing transistor. This was before integrated circuits that were just coming out. Once I wrote a program listing all the permutations of a 6 letter jumble to print out the possible words to check for the right word. It took around 120 machine instructions like shift left, add one memory location to another, print a memory location etc. Programming a machine instruction at a time is as basic as you can get.

 

[nikos79]

Once I wrote a program listing all the permutations of a 6 letter jumble to print out the possible words to check for the right word.

In Assembly?!
That's crazy.
Nowadays is just:

found = right_word in map("".join, permutations(jumble))

And found can be either True or False

 

[Alan Edward Klein]

In Assembly?!
That's crazy.
Nowadays is just:

found = right_word in map("".join, permutations(jumble))

And found can be either True or False

I don't think it was Assembly but machine language instructions. Each instruction was on a punch card.