Google begins offering free SAT practice tests powered by Gemini

hillspuck

Ars Scholae Palatinae
2,179
Of course, generative AI can go off the rails and provide incorrect information, which is a problem when you’re trying to learn things. However, Google says it has worked with education firms like The Princeton Review to ensure the AI-generated tests resemble what students will see in the real deal.

Well, that certainly allays all the fears...
 
Upvote
35 (36 / -1)

graylshaped

Ars Legatus Legionis
67,689
Subscriptor++
I'm something less than a fan of the paid test-boosting industry; the best thing I can say about this as an alternative is that it appears to be a free way to accomplish the same thing--try and shore up negligent attention to core concepts in school. I suggest the most benefit to the student comes from trying something.

And getting a good night's sleep the night before the test.
 
Upvote
11 (11 / 0)

42Kodiak42

Ars Scholae Palatinae
1,355
The best way to let young people use AI is to make their interactions with it resemble homework or studying.

If a kid asks ChatGPT about a homework question, they need to answer a bunch of Socratic questions to test their understanding first.
It's also critically important that the AI doesn't give them direct answers, but points to useable resources that contain those answers.

One of the most important skills they're going to need to learn is how to work with and contribute to reputable sources of information. That skill is only going to become more important to a person's intellectual capabilities as the internet becomes swamped with unreliable AI-generated content.
 
Upvote
21 (21 / 0)

Natrous

Seniorius Lurkius
47
Despite all the AI hate on Ars, Chat GPT and Gemini allowed me to help my high-school student daughter pass college-level calc-based physics, from the same college I graduated from years ago (and barely remembered how to do)

Many a time I was unable to figure out the homework without the help. It may feel like cheating by dropping whole questions into the prompt. But if the goal is to learn and not just to cheat, it's invaluable. It never just spits out an answer - always the steps. By asking clarification and followup questions we were able to actually understand the problems and how to attack/solve them.

She pulled a B on the hand-written final, so it certainly works.

Sure it's easy to cheat, but if you can get the proper mindset and goals into the students' heads, it's great.
 
Upvote
42 (50 / -8)

GenericAnimeBoy

Ars Tribunus Militum
1,790
Subscriptor++
Of course, generative AI can go off the rails and provide incorrect information, which is a problem when you’re trying to learn things. However, Google says it has worked with education firms like The Princeton Review to ensure the AI-generated tests resemble what students will see in the real deal.
So there's a risk of the information being incorrect, and their answer to that is that they've worked to make the incorrect information closely resemble the real test?

That's actually worse. Thanks, I hate it.
 
Upvote
5 (8 / -3)

Fred Duck

Ars Tribunus Angusticlavius
7,166
Ryan Whitwam said:
The addition of this feature to Gemini for all users will likely accelerate declines in test prep and tutoring services.
Meanwhile, the AI companies tell us that AI will definitely create jobs.

Perhaps after the fall of society?

Wouldn't it be grand if someone posted a hilarious never-before-seen illustration related to this?
 
Upvote
-3 (2 / -5)

Rector

Ars Tribunus Militum
1,543
Subscriptor++
The problem with all this is that these standardized admissions tests are coachable in the first place. It pretty much defeats the entire reason for their use.

They are coachable, but you can't coach someone from an 800 to a 1500 without them having actually learned what they are being tested on. Coaching works when the issue is mostly test-taking strategy, time management, and so on.
 
Upvote
36 (37 / -1)

marsilies

Ars Legatus Legionis
24,387
Subscriptor++
So there's a risk of the information being incorrect, and their answer to that is that they've worked to make the incorrect information closely resemble the real test?

That's actually worse. Thanks, I hate it.
Yeah, you sassed out the non-answer of their response, or rather, how the responded to the questions "is it accurate?" with "it closely resembles the actual test," instead of "the answers it gives are correct."
 
Upvote
10 (10 / 0)

42Kodiak42

Ars Scholae Palatinae
1,355
The problem with all this is that these standardized admissions tests are coachable in the first place. It pretty much defeats the entire reason for their use.
Yes and no: For most purposes, people are going to be able to prepare themselves for intellectual work they know they're going to perform and good chunk of test coaching is meant to help with the 'unique' pressures of standardized testing that do not reflect real world performance or conditions (such as that heightened time pressure, only standardized tests are designed to not give everyone enough time to answer everything they can).

The problem is that not everyone is going to have equal access to test coaching. Not only is that unfair to students who can't afford better coaching, but since coaching is paid for personally, it's probably less equal than what you're going to have access to in college and in the workplace.
 
Upvote
13 (14 / -1)

hillspuck

Ars Scholae Palatinae
2,179
Despite all the AI hate on Ars, Chat GPT and Gemini allowed me to help my high-school student daughter pass college-level calc-based physics, from the same college I graduated from years ago (and barely remembered how to do)
A lot of the AI hate on Ars derives from the ethical issues with slurping up training data without consent and the environmental effects of all these processes. That, plus whether this can actually be made to function as a business at the levels of investment it's getting or if it'll crater and depress the economy. And in the meantime, it's causing lots of supply chain issues while the datacenters monopolize hardware.

I'm glad you helped your daughter, but it doesn't really change that.
 
Upvote
27 (29 / -2)
the environmental effects of all these processes
There really isn't anything in modern life that has less of an environmental impact than AI, yet the Ars commentariat is strangely silent on those other things, just because of anti-AI vibes.

Information is good, information systems are good, machine inference systems are good, teaching kids to pass the SAT is good. If you had ever driven an ICE car across town to take your kid to an SAT test prep place, that alone did more environmental harm that a lifetime of using AI.
 
Upvote
-15 (7 / -22)

cfenton

Ars Scholae Palatinae
829
They are coachable, but you can't coach someone from an 800 to a 1500 without them having actually learned what they are being tested on. Coaching works when the issue is mostly test-taking strategy, time management, and so on.
I can't comment on the SAT specifically since I didn't do my undergrad in the US, but I did have to write the GRE to apply for grad school. In that case, getting a good score requires knowing how the test works. The math questions in particular take too long for anyone to solve in the time allotted, so you have to quickly estimate and then pick the best answer. It doesn't tell you that, and that's unlike any test I've ever written in school. Without paying for prep, you're not going to know that, and so you're going to do worse than people who do pay. It's predatory bullshit.

Don't even get me started on how these tests have very little predictive power for success in university.
 
Upvote
7 (9 / -2)

hillspuck

Ars Scholae Palatinae
2,179
There really isn't anything in modern life that has less of an environmental impact than AI, yet the Ars commentariat is strangely silent on those other things, just because of anti-AI vibes.
No, it isn't. In general, they speak out on them when there's actually an article about them.

That's how comment sections work, you know?
 
Upvote
10 (11 / -1)
No, it isn't. In general, they speak out on them when there's actually an article about them.

That's how comment sections work, you know?
I think you can see for yourself that is not the case. Go to any iPhone thread, check the comments, count up the comments that mention the GHG emissions embodied by a year of making iPhones. You will find none. Anti-AI environmental discourse is a hysteria that was 100% manufactured by the press.
 
Upvote
-18 (2 / -20)

momoisdabest

Smack-Fu Master, in training
35
A lot of the AI hate on Ars derives from the ethical issues with slurping up training data without consent and the environmental effects of all these processes. That, plus whether this can actually be made to function as a business at the levels of investment it's getting or if it'll crater and depress the economy. And in the meantime, it's causing lots of supply chain issues while the datacenters monopolize hardware.

I'm glad you helped your daughter, but it doesn't really change that.
Don't forget it encourages people to commit suicide (even those who express misgivings) and reinforces delusions. Oh it also allows for generating sexually explicit pics without consent, squeezes creatives out of their labor, allows for mass production of fake news / propaganda (see white house posting a genAI alteration of a person being arrested the other day). And it still thinks there's 2 'r' in raspberry even after hundreds of billion investments.

Now with all that said, when it does happen to generate plausible text, it's certainly cool. I see these chatbot products (incl Sora etc) more as an entertainment platform/toy than anything else
 
Upvote
9 (13 / -4)

jgee43

Ars Scholae Palatinae
702
Subscriptor++
Just leaving this here, that free SAT prep is already available from for example Khan Academy. These are officially supposed to be aligned with the exam and likely has human-examined content.

https://www.khanacademy.org/digital-sat
The Khan academy questions are definitely well-aligned. I used to teach a College Prep course, which included both ACT and SAT prep as part of the material, and the Khan Academy prep materials have been the best freely available practice for quite awhile now.

I ran some testing with Gemini just now, but it failed to incorporate/produce the graphs it referenced in a few questions. So maybe not fully functional just yet, but interesting. The questions were, in my opinion, on the weaker side compared to the real deal (especially the math), but captured the style and approach of the SAT pretty well.

It also did a good job of incorporating what I call the "sneaky" distractors that College Board loves as some of the multiple choice options. The wrong answers that you would get if you stopped calculating a step too early or if you were to mis-apply a +/- sign somewhere.
 
Upvote
11 (11 / 0)
They are coachable, but you can't coach someone from an 800 to a 1500 without them having actually learned what they are being tested on. Coaching works when the issue is mostly test-taking strategy, time management, and so on.

Yes. But when you are talking about college admissions, something I've been involved with, tens of points can make the difference between getting in or not. Well within the coachable range.
 
Last edited:
Upvote
6 (6 / 0)

hillspuck

Ars Scholae Palatinae
2,179
I think you can see for yourself that is not the case. Go to any iPhone thread, check the comments, count up the comments that mention the GHG emissions embodied by a year of making iPhones. You will find none. Anti-AI environmental discourse is a hysteria that was 100% manufactured by the press.
Remind me, when was the last time they reopened coal plants so iPhones can be made?

You're reaching so hard.
 
Upvote
9 (13 / -4)
Yes and no: For most purposes, people are going to be able to prepare themselves for intellectual work they know they're going to perform and good chunk of test coaching is meant to help with the 'unique' pressures of standardized testing that do not reflect real world performance or conditions (such as that heightened time pressure, only standardized tests are designed to not give everyone enough time to answer everything they can).

The problem is that not everyone is going to have equal access to test coaching. Not only is that unfair to students who can't afford better coaching, but since coaching is paid for personally, it's probably less equal than what you're going to have access to in college and in the workplace.

It's not about unique pressure. It's about being able to train people how to work the problems on the test. The sorts of questions asked on such tests have been shown to be coachable. It's one of the reasons a lot of schools have been trying to get away from them.

But I agree with your second paragraph. The inequity that results from access to quality test prep is the problem that results from the coachability of the tests.
 
Upvote
8 (8 / 0)
Remind me, when was the last time they reopened coal plants so iPhones can be made?

You're reaching so hard.
The fact that you are aware of that, but unaware of the carbon content of an iPhone due to coal power in Asia, is exactly the willful, curated ignorance I was referring to. I could not have made my own argument any better than you did.
 
Upvote
-9 (6 / -15)
Producing an iPhone takes about 200 kWh which is equivalent to about 60,000 chatGPT queries (assuming about 0.0003 kwh per query).
So effectively a rounding error

ChatGPT processes 2.5 billion requests a day.

edit: more calcs

This means that ChatGPT with current request numbers uses around 4000 times more energy than all of the iPhones made in a year
 
Last edited:
Upvote
12 (14 / -2)

Sadre

Ars Scholae Palatinae
1,008
Subscriptor
If the American system of education had taken a few different turns, and instead of writing essays and papers considered oral examination as the best way to test for understanding, the value, and dangers, of AI change significantly.

But we didn't go that route, and here we are, in serious trouble re: language skills. AI can do writing and thus will distort the main way we evaluate both language skills and understanding.
 
Last edited:
Upvote
0 (0 / 0)
But we didn't go that route, and here we are, in serious trouble re: language skills. AI can do writing. Speaking is another way language engages the mind.
My wife is an english as a second language teacher (for adults). She now spends half her time giving her students F grades for cheating and using ChatGPT for their assignments. It's an insanely stupid thing to do.
 
Upvote
6 (6 / 0)

Rector

Ars Tribunus Militum
1,543
Subscriptor++
My wife is an english as a second language teacher (for adults). She now spends half her time giving her students F grades for cheating and using ChatGPT for their assignments. It's an insanely stupid thing to do.
Meanwhile in Korea:

According to a Seoul National University internal survey on the 3rd, more than 97% of enrolled students use AI in their studies. It was reported that AI use has become routine in the pre-learning stages, such as writing reports, drafting papers, and preparing presentation scripts. Reflecting this trend, Seoul National University is working to build an AI system that all members can use starting in the spring semester next year.

https://biz.chosun.com/en/en-it/2025/11/03/2SQXLIX4OJE4BJFPPIR7ZGRHLI/

Why bother learning when the LLM can do all of your coursework?
 
Upvote
0 (0 / 0)

hillspuck

Ars Scholae Palatinae
2,179
The fact that you are aware of that, but unaware of the carbon content of an iPhone due to coal power in Asia, is exactly the willful, curated ignorance I was referring to. I could not have made my own argument any better than you did.
Oh, I'm aware. I'm just capable of fairly looking at the difference. The iPhone, once produced, lasts for years while fulfilling a crucial need for a phone in a modern society. As has been pointed out above, the AI queries are equivalent to producing thousands more iPhones than where made in a year, and their output in general produces nothing that lasts.
 
Upvote
8 (9 / -1)

Jensen404

Ars Scholae Palatinae
1,075
So effectively a rounding error

ChatGPT processes 2.5 billion requests a day.

edit: more calcs

This means that ChatGPT with current request numbers uses around 4000 times more energy than all of the iPhones made in a year
There are over 200 million iPhones made each year. If each phone takes as much energy to make as 60,000 ChatGPT requests, that's over 12 trillion requests worth of energy.

2.5 billion requests a day * 365 is 912.5 Billion.

So it looks like producing iPhones takes over 13 times as much energy as ChatGPT requests each year. Or am I doing something wrong here? Been over 25 years since I took an SAT.

(I'm just using the numbers given in this message thread, except looking up yearly iPhone production. It's possible they aren't taking the full energy cost of AI training and hardware into account)
 
Upvote
1 (2 / -1)

hisnyc

Smack-Fu Master, in training
82
Subscriptor
But I agree with your second paragraph. The inequity that results from access to quality test prep is the problem that results from the coachability of the tests.

I think the fact that you can prepare for the tests is actually a strength. Good test results correlate with how much additional effort you are willing to put in. That is a good indication of how you will do in college.

The problem, as the original poster indicated, is that how much effort you can put in is linked to your family's income. Before Gemini, I would have said that high scores coupled with low family income are probably an indication of motivation (you want that student!). With Gemini, it's still an indicator, but maybe not as strong a one?
 
Upvote
2 (2 / 0)

timber

Ars Scholae Palatinae
1,171
Despite all the AI hate on Ars, Chat GPT and Gemini allowed me to help my high-school student daughter pass college-level calc-based physics, from the same college I graduated from years ago (and barely remembered how to do)

Many a time I was unable to figure out the homework without the help. It may feel like cheating by dropping whole questions into the prompt. But if the goal is to learn and not just to cheat, it's invaluable. It never just spits out an answer - always the steps. By asking clarification and followup questions we were able to actually understand the problems and how to attack/solve them.

She pulled a B on the hand-written final, so it certainly works.

Sure it's easy to cheat, but if you can get the proper mindset and goals into the students' heads, it's great.
The other day I saw my kid studying for a (9th year Math) test in a similar way.
While in my time I always took care some time before to have practice materials with solutions and desirably the solving steps now he only needs the main questions.
Yes, you can cheat homework but at least currently you can't cheat tests (not more than before).
Regarding this test, I'm not from the US and my only taste of administered tests instead of regular solvable written exams (Engineering background) is the GMAT. Is it any similar? The GMAT is definitely something you should train either alone or with coaching. It probably depends a lot more in that than knowing the subject (which is mostly quite easy).
 
Last edited:
Upvote
0 (0 / 0)

Sadre

Ars Scholae Palatinae
1,008
Subscriptor
My wife is an english as a second language teacher (for adults). She now spends half her time giving her students F grades for cheating and using ChatGPT for their assignments. It's an insanely stupid thing to do.

I used "covert oral exams" my last decade+ as a professor. Term papers have been gameable for a long time. But if I am using your paper meetings and draft meetings as a ready-made format for oral examination (it is NOT hard to do, I was taught to do it), that paper better match up with my copious notes from our meetings.

(Had class sizes <20 students. Oral examination is time consuming. I would be on campus sometimes till midnight. I just love those paper meetings!)

Could not do oral examination openly b/c in US oral exams are (were?) viewed as not good. "Make them write!" was the rule.

Americans hate to write by and large. But you can ask them questions face to face and that's that.
 
Upvote
2 (2 / 0)