Grammarly's AI Detection: Bias and Process
Key Points
- Grammarly’s new “Authorship” feature aims to flag AI‑generated text, but its methodology—detecting words and patterns more common in AI output—raises major accuracy concerns.
- The system is likely to be biased against non‑native English speakers, whose distinctive word‑choice patterns can trigger false AI detections.
- Emphasizing process‑based policing (labeling text as AI‑generated) over outcomes distracts from teaching students how to use AI as a productive tool.
- Because AI authorship cannot be proven deterministically, the tool risks generating false positives that could be misused to make unsupported accusations.
- While plagiarism detection (copy‑and‑paste from academic sources) could be valuable, the broader “authorship” enforcement may hinder learners from developing critical, outcome‑focused AI skills.
Sections
Full Transcript
# Grammarly's AI Detection: Bias and Process **Source:** [https://www.youtube.com/watch?v=8MT0xlArmmY](https://www.youtube.com/watch?v=8MT0xlArmmY) **Duration:** 00:06:36 ## Summary - Grammarly’s new “Authorship” feature aims to flag AI‑generated text, but its methodology—detecting words and patterns more common in AI output—raises major accuracy concerns. - The system is likely to be biased against non‑native English speakers, whose distinctive word‑choice patterns can trigger false AI detections. - Emphasizing process‑based policing (labeling text as AI‑generated) over outcomes distracts from teaching students how to use AI as a productive tool. - Because AI authorship cannot be proven deterministically, the tool risks generating false positives that could be misused to make unsupported accusations. - While plagiarism detection (copy‑and‑paste from academic sources) could be valuable, the broader “authorship” enforcement may hinder learners from developing critical, outcome‑focused AI skills. ## Sections - [00:00:00](https://www.youtube.com/watch?v=8MT0xlArmmY&t=0s) **Untitled Section** - ## Full Transcript
I want to talk about grammarly and I
want to talk about why we use tools for
process instead of tools for outcome and
how we can't really detect AI so
grammarly announced that they are going
to roll out a product called authorship
that is supposed to detect AI if you
read the fine print on grammarly's own
page there are a number of massive
massive concerns around this roll out
they admit it could be biased against
non-native English speakers they admit
that this is based on detecting words
that AI uses more frequently they're
using an algorithm they're reading for a
pattern in text that they think AI uses
more frequently but of course non-native
English speakers could also use that
pattern because they don't necessarily
have the same underlying distribution of
word choice there's often other grammars
that that they're working with and I say
that as an Indonesian speaker and I see
it
myself when I look at that the first
thing I think of is why are you focused
so much on the process instead of on the
outcome and that is probably me stuck in
tech for too long because I understand
that if you're a student there is a lot
of concern around authorship and Ai and
grammarly is trying to be a good company
and ship a product that addresses
that but I am concerned that if they
ship that product it is going to be used
to generate false positives it's going
to be used to make claims that can't
really be backed up by a deterministic
check you cannot know if the text was
really a I generate it yes you can say
oh it uses delve it uses interplay it
uses a generic
adjective sure so do
people I I and one of the challenges
that that you have to face is that
you're working in a medium that people
use and it could well be just the
idiosyncrasies of someone's language use
that you are falsely dinging now where I
think that the product as value is they
say that they're checking for word for
word copies in a lot of academic
journals and they're looking to see if
there's plagiarism where people are
copying text from other papers that
maybe the professor hasn't read and
putting it into the document that that
could really help right that could help
to reduce the incidence of
plagiarism I do not think that that is
that that the authorship tool more
broadly is actually going to help with
the problem they are professing to solve
around authorship detection I don't
think so I also don't think it's worth
it because at the end of the day from a
larger economic perspective you need
these students to be prepared to use AI
in the workplace and if you are focused
exclusively on process on for instance
and I'm actually not kidding here
literally labeling your Google doc text
with what is what what the system thinks
is AI generated and what it isn't you
are teaching the students to Wordsmith
and actually lose time on AI instead of
teaching them to use AI to deliver
effective outcomes think critically
about whether the AI is getting them
there and go faster and those are the
skills that students are going to need
in business those are the schools that
professionals around me are using all
the time is it worth it is this the
right call for AI it's it's intelligence
allocation do I give this to the llm to
do or do I do it and grammarly actually
isn't supporting that motion when they
do this they are very much inserting
themselves into the debate between
profess and students and while that may
make sense for their customer base I'm
not inside grammarly I don't know it is
something that reflects a
shortsightedness with regards to
students long-term career preparation
for the workplace students are going to
need to know how to use large language
models at
work otherwise they're going to be
caught without that skill set and
they're going to be less
competitive it does not make sense to me
to try and make them put the breaks on
that when we should actually be asking
them to think critically about how
they're using it and to make sure that
the outcom that they're delivering are
better and so I would like to see are
llm papers better are papers that maybe
maybe even that that this tool thinks
are AI generated are they better is the
outcome better because if the quality
bar is being
raised and you still want to spend time
labeling and wordsmithing you're
labeling and wordsmithing so that the
quality bar on average can come down to
match what it used to be because now it
will be human generated and that does
not make sense to me that is not a
motion that anybody in business is going
to
support and I'm not saying that because
education should just be like a pipeline
to business but I am calling out that
there are career skills here in a very
fast changing world that matter and I
feel like this is a product release that
does not support long-term Career
Development and long-term skill growth
for people who are getting started in a
world where llms are going to be there
from day one of their careers they have
to know how to use them
well but I have not heard of anybody at
the University level that is focused on
making sure that they understand how
llms Drive business
outcomes nobody now I've heard you know
this is how you use chat GPT yes
absolutely that's getting taught but I
would venture to Guess that if you think
that you can use an AI authorship tool
to reliably detect student text versus
AI text then you're not teaching how to
use chat GPT well because you know what
you can do you can go into the system
prompt in chat GPT and you can tell it
don't use generic adjectives don't use
delve don't use interplay you can name
whatever you want and it
won't and now what are you going to do
with your AI detected
text so you can hack around the system
already and if they don't realize the
students are doing it they're probably
not teaching the students well enough so
I call all of this out because one one
there is a persistent myth that AI
detection is deterministic and you can
just know and it's not true and two I
call it out because I think we need to
think more about outcomes and how we
deliver extra value with the AI we
generate and the time we save and I
think we need to think less about
whether we need to make our writing
sound a particular way I don't think
that's very productive at all so that's
my thought on grammarly and authorship
don't love that release understand why
they did it I think the best thing about
it is the plagiar plagiarism detector
from other academic papers what do you
think