have decided that the only way they can find out what the chatbots will do in the real world is by letting them loose — and reeling them in when they stray. They believe their big, public experiment is worth the risk.

In the days since the Bing bot’s behavior became a worldwide sensation, people have struggled to understand the oddity of this new creation. More often than not, scientists have said humans deserve much of the blame.

“Whatever you are looking for — whatever you desire — they will provide.”

But there is still a bit of mystery about what the new chatbot can do — and why it would do it. Its complexity makes it hard to dissect and even harder to predict, and researchers are looking at it through a philosophic lens as well as the hard code of computer science.

Why Do A.I. Chatbots Tell Lies and Act Weird? Look in the Mirror.

critique the products of AI writing tools

an intimidating blinking cursor on a blank page

We should be familiarizing ourselves with, and nurturing, our student’s writing styles and lines of inquiry.

In this moment of generative AI, Hypothesis continues to rely on what we’ve always done: support process-oriented pedagogies that make learning more accessible.

It’s hard to avoid concerns about plagiarism with the rise of ChatGPT.

A calculator performs calculations; ChatGPT guesses. The difference is important.

synthetic writing

In PRS, I encouraged teachers to shift their focus from asking questions to teaching students how to ask high-quality questions themselves.

However, the article does not take a clear stance on the matter and does not offer a conclusion on whether the capitalization of the word "black" is a good or bad thing.

long overdue

John Eligon, a New York Times columnist, writes about the debate surrounding the capitalization of the word "black" to refer to African-Americans. The move to capitalize the word has gained widespread acceptance in recent weeks and has sparked a debate over identity, race, and power. Many news organizations have changed their style to capitalize the word, but not all have embraced it, including The New York Times and The Washington Post, who are still grappling with the decision. The capitalization of black is part of a long-standing struggle to refer to people of African descent in the most respectful and accurate way.

upskilling activities in areas like writing and coding (debugging code, revising writing, asking for explanations)

ChatGPT use in Bibtex format as shown below:

will continue to broaden as we learn.

As part of this effort, we invite educators and others to share any feedback they have on our feedback form as well as any resources that they are developing or have found helpful (e.g. course guidelines, honor code and policy updates, interactive tools, AI literacy programs, etc).

one factor out of many when used as a part of an investigation determining a piece of content’s source and making a holistic assessment of academic dishonesty or plagiarism.

Ultimately, we believe it will be necessary for students to learn how to navigate a world where tools like ChatGPT are commonplace. This includes potentially learning new kinds of skills, like how to effectively use a language model, as well as about the general limitations and failure modes that these models exhibit.

Some of this is STEM education, but much of it also draws on students’ understanding of ethics, media literacy, ability to verify information from different sources, and other skills from the arts, social sciences, and humanities.

While tools like ChatGPT can often generate answers that sound reasonable, they can not be relied upon to be accurate consistently or across every domain. Sometimes the model will offer an argument that doesn't make sense or is wrong. Other times it may fabricate source names, direct quotations, citations, and other details. Additionally, across some topics the model may distort the truth – for example, by asserting there is one answer when there isn't or by misrepresenting the relative strength of two opposing arguments.

While the model may appear to give confident and reasonable sounding answers,

. Web browsing capabilities and improving factual accuracy are an open research area that you can learn more in our blog post on WebGPT.

subtle ways.

Verifying AI recommendations often requires a high degree of expertise,

students may need to develop more skepticism of information sources, given the potential for AI to assist in the spread of inaccurate content.

Educators should also disclose the use of ChatGPT in generating learning materials, and ask students to do so when they incorporate the use of ChatGPT in assignments or activities.

export their ChatGPT use and share it with educators. Currently students can do this with third-party browser extensions.

they and their educators should understand the limitations of the tools outlined below.

custom tutoring tools

Scaling neural network models—making them bigger—has made their faux writing more and more authoritative-sounding, but not more and more truthful.

Why are people so quick to be impressed by the output of large language models (LLMs)?

So will AI text generation tools revolutionize or kill college writing? Both! Neither! For sure! Probably! Eventually! Somewhat! It’s…complicated.

e-Literate isn’t about what I know. It’s about what I’m learning.

Particularly if used judiciously as part of the writing curriculum rather than the whole thing, it could be quite useful.

students are heavily influenced by whether they believe their teacher cares about their learning.

Then I would have edited the output

ChatGPT doesn’t mark the end of high school English class, but it can mark the end of formulaic, mediocre writing performance as a goal for students and teachers. That end is long overdue, and if ChatGPT hastens that end, then that is good news.

Both started with a version of “Work A and Work B have many similarities and many differences,” an opening sentence that I would have rejected from a live student

If they put a great degree of thought into designing a prompt, would that not mean that they were doing something involving real learning?

I suspect that test runs with ChatGPT depend in part on the richness of the prompt given,

And the algorithm cannot manage supporting its points with quotes from the works, a pretty fundamental part of writing about literature.

He said it was “very naive” to think it would be possible to impose restrictions on internet platforms, particularly with Microsoft primed to integrate AI into its search engine, Bing.“Are you going to ban Google and Bing?”

Ready to take your creativity to the next level? Look no further than generative AI!

AI is increasingly used for assessments of student learning. These are all valuable and enhance our efficiency and effectiveness.

At the same time, we need to continue building activities and assessments to make classroom work more specific and experiential.

Pedagogically speaking, focusing on the grunt work of trying out ideas—watching them develop, wither, and cede ground to better ones—is the most valuable time we can spend with our students. We surrender that time to Silicon Valley and the messy database that is the internet at the peril of our students.

digital utopians might claim that students and teachers will have more opportunities for critical thinking because generating ideas—the grunt work of writing—isn’t taking up any of our time. Along this line of thinking, ChatGPT is just another calculator, but for language instead of numerical calculation.

Analysis of recent events not in the training data for the system.

Note that ChatGPT can produce outputs that take the form of  “brainstorms,” outlines, and drafts. It can also provide commentary in the style of peer review or self-analysis. Nonetheless, students would need to coordinate multiple submissions of automated work in order to complete this type of assignment with a text generator.

get a better sense of their thinking

a process that empowers critical thinking

students who feel connected to their writing will be less interested in outsourcing their work to an automated process.

skip the learning and thinking around which their writing assignments are designed.

Augmenting teachers, not replacing them

There’s a line somewhere between using ChatGPT in collaboration, and getting it to do all the work.

ChatGPT is not an original thinker, but you are.

using the model’s suggestions as a starting point

Right now, one of the most powerful things you can learn about ChatGPT is how to write quality prompts.

Beyond the media hype about cheating,

"I would much rather have ChatGPT teach me about something than go read a textbook."

the company has also heard from them that the chat bot can be "an unbelievable personal tutor for each kid," Altman said.

"We adapted to calculators and changed what we tested for in math class, I imagine.

, an arithmetic operator in Python is not a function. An arithmetic operator is a symbol that performs

This framing means that as educators we need to be clear not only about what we hope our students are learning but also about how and why.

The main goal of transparent teaching is simple: to promote students’ conscious understanding of how they learn.

The educational issues surrounding ChatGPT are similar in kind to those we've seen with the growing power of the web

Note that students will not be able to cite ChatGPT using a link to their generated response;instead, ask students to repeat the exact language of their search query in the footnotes in lieu of a link

formulaic syntax

These tools, along with a range of other practices,

scrape any website using tab gbt

I've been using ChatGPT pretty consistently during the workday and have found it useful for open ended programming questions, "cleaning up" rough bullet points into a coherent paragraph of text, etc. $20/month useful is questionable though, especially with all the filters. My "in between" solution has been to configure BetterTouchTool (Mac App) with a hotkey for "Transform & Replace Selection with Javascript". This is intended for doing text transforms, but putting an API call instead seems to work fine. I highlight some text, usually just an open ended "prompt" I typed in the IDE, or Notes app, or an email body, hit the hotkey, and ~1s later it adds the answer underneath. This works...surprisingly well. It feels almost native to the OS. And it's cheaper than $20/month, assuming you aren't feeding it massive documents worth of text or expecting paragraphs in response. I've been averaging like 2-10c a day, depending on use.Here is the javascript if anyone wants to do something similar. I don't know JS really, so I'm sure it could be improved. But it seems to work fine. You can add your own hard coded prompt if you want even. async (clipboardContentString) => { try { const response = await fetch("https://api.openai.com/v1/completions", { method: "POST", headers: { "Content-Type": "application/json", "Authorization": "Bearer YOUR API KEY HERE" }, body: JSON.stringify({ model: "text-davinci-003", prompt: `${clipboardContentString}.`, temperature: 0, max_tokens: 256 }) }); const data = await response.json(); const text = data.choices[0].text; return `${clipboardContentString} ${text}`; } catch (error) { return "Error" } }

create assessments that “take into consideration the processes and experiences of learning.”

Ask students to engage in metacognitive reflection that has them articulate what they have learned, how they have learned it, and why the knowledge is valuable.

Is this moment more like the invention of the calculator, saving me from the tedium of long division, or more like the invention of the player piano, robbing us of what can be communicated only through human emotion?

The question isn’t “How will we get around this?” but rather “Is this still worth doing?”

The rudiments of writing will be considered a given, and every student will have direct access to the finer aspects of the enterprise.

Last night, I received an essay draft from a student. I passed it along to OpenAI’s bots. “Can you fix this essay up and make it better?” Turns out, it could. It kept the student’s words intact but employed them more gracefully; it removed the clutter so the ideas were able to shine through. It was like magic.

nor does it successfully integrate quotations from the original texts

What GPT can produce right now is better than the large majority of writing seen by your average teacher or professor.

methods of assessment that take into consideration the processes and experiences of learning, rather than simply relying on a single artifact like an essay or exam. The evidence of learning comes in a little of different packages

The fact that the AI writes in fully fluent, error-free English with clear structure virtually guarantees it a high score on an AP exam

ChatGPT may be a threat to some of the things students are asked to do in school contexts, but it is not a threat to anything truly important when it comes to student learning.

an opportunity to re-examine our practices and make sure how and what we teach is in line with our purported pedagogical values.

Rather than letting students explore the messy and fraught process of learning how to write, we have instead incentivized them to behave like algorithms, creating simulations that pass surface-level muster

The meaning of life is

In The New Laws of Robotics, legal scholar Frank Pasquale argues for guidance from professional organizations about whether and how to use data-driven statistical models in domains such as education or health care.

we need collaborative processes to seek clarity.

slow thinking,

Should I ask students to prompt a language model and then critique its output?

preferences of data scraped from internet sites hardly renowned for their wisdom or objectivity.

“mathy math,” a model of language sequences built by “scraping” the internet and then, with massive computing, “training” the model to predict the sequence of words most likely to follow a user’s prompt

What a contrast to the masochistic persistence I had practiced for so many years and preached to my struggling students.

Back in the early 2000s, I used to demonstrate to students how EasyBib often gets it wrong when it comes to MLA formatting.

having students socially annotate the paper, practicing their editing and fact-checking skills.

The text is being generated on behalf of the student and is being substituted for the student’s self-generated text. This use of AI is inherently dishonest.

Teaching students to be good critical readers takes time and requires instructors develop activities, such as social annotation assignments, that draw students’ attention to the details of a well-written text.

So, while effective editors may or may not be exceptional writers, they must be great critical readers.

Pedagogy: Some Ideas on How to Use GPT in Teaching

I could instead present students with ChatGPT’s response alongside some marking instructions and ask them to provide a critique on what grade the automated response deserves and why.

do we really need all students to be writing the same essays and responding to the same questions?

an opportunity to improve the way we assess

articulate its inability to fully replicate the expertise and real-world experience that human teachers bring to the classroom

If ChatGPT is used to grade assignments or exams,

making it capable of engaging in natural language conversations.

Figure 3. The average drop in log probability (perturbation discrep-ancy) after rephrasing a passage is consistently higher for model-generated passages than for human-written passages. Each plotshows the distribution of the perturbation discrepancy d (x, pθ , q)for human-written news articles and machine-generated arti-cles; of equal word length from models GPT-2 (1.5B), GPT-Neo-2.7B (Black et al., 2021), GPT-J (6B; Wang & Komatsuzaki (2021))and GPT-NeoX (20B; Black et al. (2022)). Human-written arti-cles are a sample of 500 XSum articles; machine-generated textis generated by prompting each model with the first 30 tokens ofeach XSum article, sampling from the raw conditional distribution.Discrepancies are estimated with 100 T5-3B samples.

if we apply small perturbations to a passagex ∼ pθ , producing ̃x, the quantity log pθ (x) − log pθ ( ̃x)should be relatively large on average for machine-generatedsamples compared to human-written text.

As in prior work, we study a ‘white box’ setting (Gehrmannet al., 2019) in which the detector may evaluate the log prob-ability of a sample log pθ (x). The white box setting doesnot assume access to the model architecture or parameters.While most public APIs for LLMs (such as GPT-3) enablescoring text, some exceptions exist

Empirically, we find predictive entropy to be positively cor-related with passage fake-ness more often that not; there-fore, this baseline uses high average entropy in the model’spredictive distribution as a signal that a passage is machine-generated.

We find that supervised detectors can provide similardetection performance to DetectGPT on in-distribution datalike English news, but perform significantly worse than zero-shot methods in the case of English scientific writing andfail altogether for German writing. T

ex-tending DetectGPT to use ensembles of models for scoring,rather than a single model, may improve detection in theblack box setting

hile in this work, we use off-the-shelfmask-filling models such as T5 and mT5 (for non-Englishlanguages), some domains may see reduced performanceif existing mask-filling models do not well represent thespace of meaningful rephrases, reducing the quality of thecurvature estimate.

For models be-hind APIs that do provide probabilities (such as GPT-3),evaluating probabilities nonetheless costs money.

We simulate human re-vision by replacing 5 word spans of the text with samplesfrom T5-3B until r% of the text has been replaced, andreport performance as r varies.

Figure 5. We simulate human edits to machine-generated text byreplacing varying fractions of model samples with T5-3B gener-ated text (masking out random five word spans until r% of text ismasked to simulate human edits to machine-generated text). Thefour top-performing methods all generally degrade in performancewith heavier revision, but DetectGPT is consistently most accurate.Experiment is conducted on the XSum dataset

DetectGPT’s performancein particular is mostly unaffected by the change in languagefrom English to Germa

ecause the GPT-3 API does not provideaccess to the complete conditional distribution for each to-ken, we cannot compare to the rank, log rank, and entropy-based prior methods

improving detection offake news articles generated by 20B parameterGPT-NeoX

his approach, which we call DetectGPT,does not require training a separate classifier, col-lecting a dataset of real or generated passages, orexplicitly watermarking generated text. It usesonly log probabilities computed by the model ofinterest and random perturbations of the passagefrom another generic pre-trained language model(e.g, T5)

See ericmitchell.ai/detectgptfor code, data, and other project information.

The real danger is not to people who are experts in their fields. Super experts in every field will continue to do what they have always done. All of us, however, are novices in almost everything we do. Most of us will never be experts in anything. The vast majority of the human experience of learning about something is done at the novice level. That experience is about to be autotuned.

Starting this year, we’re going to be returned a mishmash of all the information that is available on the Internet, sorted by mysterious practices

but an opinion is different from a grounded understanding.

Humanities departments judge their undergraduate students on the basis of their essays. They give Ph.D.s on the basis of a dissertation’s composition. What happens when both processes can be significantly automated?

It can summarize things you’ve said to it in new language that helps you look at yourself in a different light and reframe situations more effectively. 

Journaling in GPT-3 feels more like a conversation, so you don’t have to stare at a blank page or feel silly because you don’t know what to say.

If you know how to use it correctly and you want to use it for this purpose, GPT-3 is pretty close, in a lot of ways, to being at the level of an empathic friend

more and more jobs involve the use of generative AI for everything from discovering new drug molecules to developing ad copy,

more creative assessments that require students to demonstrate application of knowledge rather than simply the ability to produce information.

It forces us to reconsider what is distinctly human about intelligence if a machine can generate human language complete with analysis.

I fully believe that the fact that the essay was written by AI and not a live person would be undetectable for many college admissions committees.

synthesize

deeper humanistic questions like, what is truth? What is. beauty? How do we know what we know?

I'm happy to say good riddance to the college essay and other "skills" that we've come to see as proving the value of the humanities.

one that demands students find out something about themselves and tell it to you in a voice that is their own.

One high school teacher told me that he used ChatGPT to evaluate a few of his students’ papers, and that the app had provided more detailed and useful feedback on them than he would have, in a tiny fraction of the time.

They’ll need to know their way around these tools — their strengths and weaknesses, their hallmarks and blind spots — in order to work alongside them. To be good citizens,

If my goal as a teacher is to help students learn, then should I withhold access to the same information until the timing is convenient for me? Some teachers might look at this and ask if they’re even relevant anymore. I would suggest that the answer is yes, but our focus will shift on helping students develop even deeper critical thinking skills in English. Less motivated teachers who just want to teach grammar may be weeded out over time, but those who are ready to take deeper dives will get the opportunity to mentor students on their path to becoming confident and autonomous English users.

ChatGPT can be used as a writing assistance tool. ELLs can use ChatGPT to generate ideas and receive feedback on their writing. ChatGPT can also provide grammar and spelling assistance, which can be particularly helpful for ELLs who are still learning the rules of the English language.

you can use pretty much any device with a microphone to transcribe your spoken English into the prompt box. Once the responses are generated, students can use screen reading software to verbalize the response.

In OOP, objects are used to represent real-world entities, and the methods (i.e., functions) and attributes (i.e., v

At the end of the day, Copilot is supposed to be a tool to help developers write code faster, while ChatGPT is a general purpose chatbot, yet it still can streamline the development process, but GitHub Copilot wins hands down when the task is coding focused!