0:01 So by the end of this video, you'll have
0:04 full understanding of how AI detectors
0:07 work. So what exact elements of our
0:09 writing these AI detectors pay attention
0:12 to. So if you've been using Chajbd or
0:13 some other AI tool to support your
0:15 writing or to write some paragraphs for
0:17 you now, you don't want to be flagged
0:20 for uh AI generated content, you want to
0:22 know exactly how to redo these
0:24 paragraphs. And let's be honest, you
0:26 don't want to be using these AI
0:28 humanizer tools because they are
0:30 completely useless. they will uh heavily
0:32 affect the quality of your writing. So
0:34 you have to know how to work on this
0:36 yourself and how to restructure and then
0:39 I'll show you an example of AI generated
0:42 paragraph and I'll work on it and try to
0:43 read it so that it doesn't get flagged
0:46 for AI generated content. So first let's
0:48 quickly go through uh the different
0:49 factors or different elements of your
0:52 writing that uh AI detectors uh pay
0:54 attention to. So it's not just about the
0:56 words as some people think. It's not
0:59 just about these favorite words that AI
1:02 uses. It's about so much more than that.
1:04 So the first metric and these two are
1:06 the most commonly cited statistical
1:09 metrics or text perplexity and
1:11 burstiness. So basically perplexity is a
1:13 measure of how surprised a language
1:15 model is by our text because human
1:17 writing tends to have a higher
1:19 perplexity because it's more
1:21 unpredictable. AI content on the other
1:24 hand is more statistically uh expected
1:26 and therefore has less perplexity and
1:30 burstiness is pretty related is is uh
1:32 about variation. So so human writing
1:35 again it it has variation in sentence
1:39 length structure word choice while uh AI
1:41 tends to be just uniform. Then sentence
1:44 structure and grammar. So consistency
1:47 and polish these are things that are
1:49 taken into account uh because AI
1:51 generated text is usually grammatically
1:54 flawless uh and stylistically
1:56 consistent. Quite often it's just too
1:58 consistent. Then AI does have its
2:00 favorite words like I said it may repeat
2:03 words, phrases or sentences and it just
2:05 does it uh too often. And another one is
2:07 very subtle. It's about transitions.
2:09 It's good to use these expressions, but
2:11 AI likes to overuse these expressions,
2:12 these connectors like therefore,
2:15 however, in conclusion. Again, this
2:16 overlaps with the the favorite
2:18 expressions that it just like to to
2:21 overuse. Vocabulary and word choice very
2:23 related again overlaps. It's about
2:24 favorite expressions. At the same time,
2:28 it's using safe vocabulary. So, very
2:30 rarely does it use any colloquial words.
2:33 Um, for our academic writing is not a
2:35 big problem. So here I just chose
2:37 whatever is relevant for academic
2:39 writing of course and then when we talk
2:40 about fixing things again I'm talking
2:42 about fixing for academic writing
2:44 because you can see some guidelines
2:46 online for how to fix how to humanize
2:48 content and we will talk about things
2:50 like adding humor personal anecdotes
2:52 making it inconsistent and open-ended
2:54 but it's usually talking about our
2:56 everyday language. So maybe if you're
2:58 writing a blog here, we're talking about
3:01 academic text of course. So it so the
3:03 bar is much higher. But about the
3:05 vocabulary and AI, it's learned to
3:07 predict what comes next. So again, it
3:09 takes this most predictable path
3:12 essentially when writing our content for
3:14 us and then later gets our content gets
3:16 flagged by AI for the exact same reason.
3:19 Terms of coherence and depth of AI
3:21 writing. Again, it's not always easy to
3:24 spot, but it does focus on the surface
3:27 uh level meaning quite often lacks
3:29 originality argumentation. And then it
3:31 uses the statements that do seem fine
3:33 for academic writing, but again it
3:34 really depends on the context of where
3:36 they appear. So they appear meaningful,
3:37 but they are not always meaningful. So
3:39 like it's important to consider all
3:41 perspectives. It can be a meaningful
3:43 statement or it can be just there for
3:45 the sake of it. And this often the
3:47 latter often happens in AI writing. And
3:50 then finally it may compare our writing
3:52 to know other known outputs. So their
3:55 language corpora and uh both for AI
3:56 language and human language and some of
4:00 these detectors they make use of this uh
4:02 huge database of language of any
4:04 possible use of language that there
4:06 exists. So so again that's another
4:08 factor. So you can see that there are
4:10 more factors than just the words that AI
4:12 uses. This is what people tend to think.
4:13 So they think that they will just
4:16 replace these favorite expressions of uh
4:19 AI or CHBD and the text will become
4:20 undetectable and then they're surprised
4:23 that it's still uh very much detectable.
4:25 So it's there is so much more. It's
4:26 mainly if I had to pick one, it is
4:28 really about the structure. You have to
4:30 introduce variation in the structure.
4:32 We'll go through it. So now I'll go
4:34 through some ideas for how to humanize
4:36 that text. and I I'll continue to talk
4:40 about some issues with AI uh created or
4:42 generated content and then I'll show you
4:44 an example of AI generated content how
4:46 it's flagged 100% AI generated and then
4:49 how I'll work on some uh some changes
4:52 and to make it less or to make it
4:54 undetectable. So the first way to fix AI
4:56 generated content and humanize it is to
4:58 introduce intellectual hesitation which
5:00 is a normal thing that is expected in
5:03 academic writing anyway. So basically AI
5:05 likes to write in absolutes in uh
5:07 present information as factual
5:09 statements as something that's just sure
5:11 and scholars on the other hand or humans
5:13 for that matter they hedge they suggest
5:16 they suspect. So so it's about using
5:18 this kind of language which again in any
5:20 case is something that you have to do in
5:22 academic writing. So you don't like you
5:24 don't want to be presenting things as
5:26 facts. Instead you're talking about uh
5:28 there uh being a possibility that
5:30 something happens. it appears that
5:33 something is true. Uh it is suspected,
5:35 it is believed, it is likely. So this
5:37 kind of thing, this is something that AI
5:39 rarely rarely does. So it will say
5:42 social media has an impact on uh
5:44 whatever self-esteem and it's not
5:46 something that we like to do like I said
5:48 in any case, but especially if we want
5:50 to avoid being flagged for AI content.
5:52 Another one is to add subtle critique.
5:55 Critique critique critique
6:00 critique critique. Jesus. A subtle crit
6:04 critique critique critique
6:11 critique
6:13 pronunciation critique. Another one is
6:15 to add subtle critique. And I'm laughing
6:17 because I just spent a little bit of
6:19 time trying to trying to learn how to
6:21 pronounce the word the word critique.
6:22 This is again pretty obvious because in
6:24 academic writing you do have to present
6:27 different uh points of view. Anyway, so
6:30 AI will present things uh it will just
6:32 factually present a statement that
6:35 somebody does or says something whereas
6:37 as a human we want to point out maybe
6:39 some inconsistencies or point to some
6:41 other points of view. It's not something
6:42 you can do to every sentence. So we'll
6:43 go through every sentence in a second
6:45 but it's something that generally should
6:47 be present. And now this one is very
6:49 important. This one will be uh important
6:51 at the level of sentences and and just
6:54 has to be uh present everywhere as we
6:57 make our way through this AI generated
7:00 context. So to introduce uh variation in
7:03 sentence opening uh openings and syntax.
7:06 So AI will often start sentences in the
7:08 exact same way even if we don't
7:10 initially notice it. So things like this
7:13 study this shows it is important. So
7:15 there are these little repetitions and
7:17 again they are very uh to our basically
7:19 human eye it's it's very difficult to
7:22 spot but they are almost exact uh in
7:24 their structure and their length which
7:25 then adds to the rest of the structure
7:27 of the sentence and overall the whole
7:29 sentence tends to have the same kind of
7:31 number of beats same number of sometimes
7:33 even similar number of syllables words
7:35 length and so on and so forth. So by
7:38 adding this this variation and how we
7:40 introduce our sentences we are breaking
7:42 down that structure already. So we can
7:44 use introductory clauses, dependent
7:46 clauses, inverted structures rather than
7:49 just uh repeating that same exact
7:50 structure of a sentence. So things like
7:53 although widely cited, this study has
7:54 received little empirical follow-up.
7:56 It's already a little bit different or
7:57 central to this argument is the notion
8:00 of so just uh reversing the order.
8:01 You'll see when I go through individual
8:03 uh sentences what I mean just reversing
8:05 the order and introducing this
8:06 variation. So now you know what factors
8:09 are being considered by AI detectors.
8:12 how you know roughly in theory how to
8:14 humanize the content. And now let's have
8:16 a look at how we can actually humanize
8:18 the content. So let's have a look at
8:21 exact example sentence generated by AI
8:23 and see how we can change it and
8:25 humanize it so it's undetectable. But
8:26 before I continue, have a look at my
8:28 website and the range of services that I
8:30 offer that are relevant to any aspect
8:32 and any stage of your study. They
8:34 include Zoom tutorials, data analysis
8:36 services or writing support. Okay. So,
8:38 what we can see here now is a paragraph
8:42 that I asked Chad GBT to produce. If the
8:43 sound quality is slightly different,
8:45 it's just because I switched computers.
8:46 Hopefully, it's not too bad. But this is
8:48 the paragraph that we're going to to
8:51 work on. Now, just to show you that it
8:53 it is being flagged for AI generated
8:55 content. Let me paste it here to this
8:59 free tool, detect AI. Uh, sure enough,
9:02 100% generated by AI. So, we'll have to
9:04 work on it a little bit. So I pasted it
9:07 to Microsoft Word and let uh let us
9:08 begin the work. So the first sentence,
9:10 self-esteem plays a critical role in
9:12 shaping a communicative experience of
9:14 migrants using English second language.
9:15 Firstly, if I liked it, I could just
9:17 keep it. Uh as I said, what is really
9:19 more important are not individual
9:21 sentences, just how they are glued
9:22 together. So uh occasionally, if I
9:25 really like something, I can change it.
9:26 uh as long as I make sure that the
9:29 following sentence does not follow the
9:31 exact same pattern like I said the
9:34 number of beats and and uh generally the
9:37 length. So uh here I will uh tone it
9:39 down a little bit. So I'll say as I said
9:41 uh I don't like and nobody does in
9:44 academia to present things as factual
9:45 statements. So rather than it plays a
9:49 critical role, I'll say it uh can play a
9:51 critical role. And at the end, we'll
9:52 just add the reference because again,
9:55 it's something that normally would be
9:57 expected to be referenced anyway. So,
9:59 let me just say it's some somebody
10:03 called Jackson 2020. And this already
10:04 kind of breaks down the sentence a
10:06 little bit. And the second sentence is
10:08 high self-esteem fosters confidence,
10:10 which is essential for engaging in
10:12 conversations, expressing needs, and
10:14 participating social, educational, and
10:17 professional context.
10:20 I don't like some of it. For example,
10:22 participating in social educational
10:25 professional context seems a little bit
10:27 vague. So, so this is another in my
10:29 opinion a little bit of a giveaway. Not
10:31 so much, but it's just something I don't
10:35 like. In any case, uh it's talking about
10:37 uh confidence, which is essential for
10:39 engaging conversations. So, we're
10:40 talking about conversations, which is
10:42 fine. So, why then talk about social and
10:44 participating in social education
10:46 professional context? So I feel like we
10:48 just have to simplify it. So that's
10:49 generally my approach. I have to
10:51 understand the sentence first and then
10:54 think of how I can possibly uh talk
10:56 about it in a way that I just like more
10:59 than what it says now. And usually this
11:01 includes deleting some things that are
11:02 just not needed. Remember what I told
11:04 you about surface level meaning and
11:06 things it says that kind of make sense
11:08 but then they don't really. So I'll
11:10 start by again toning it down. I'll say
11:16 it um has also been shown that high
11:19 self-esteem can increase so so again
11:23 toning down can increase uh
11:24 uh
11:26 confidence which is a chance for
11:28 engaging blah blah blah and then
11:30 participating there is just too much
11:32 information can increase
11:36 confidence and as a result um their
11:38 willingness or migrants doesn't matter
11:42 their willing ness to to
11:44 participate in various
11:47 various
11:50 conversational contexts exchanges. It
11:52 doesn't matter. It's not perfect either,
11:54 but basically the the
11:57 the message is the same. Hopefully, you
11:59 agree there's too much about the
12:01 conversation and expressing needs. It's
12:02 just about confidence and being willing
12:04 to participate, to engage in
12:06 conversation. So, now let's continue.
12:08 Again, it's not a bad word and I use it
12:11 conversely. I just feel like here I'll
12:14 include some variation again to to
12:16 influence the overall length even of the
12:19 sentence. I'll say low self-esteem I'll
12:22 also uh reverse the order as I said
12:24 earlier. So low low self-esteem on the
12:26 other hand and there is a lot to unpack
12:28 here again. So it may hinder
12:30 communication but by increasing anxiety,
12:34 fear of judgment and reluctance to speak
12:36 which doesn't make that much sense
12:38 because again it's uh or maybe it makes
12:40 some sense anyway. Ultimately affecting
12:42 language development and social
12:44 integration which is again a little bit
12:45 about communication. It's just about
12:48 simplifying. I want to simplify this. Uh
12:52 so low self-esteem on the other hand may
12:56 affect social integration mainly due to
13:00 increased anxiety and fear of judgment.
13:02 It doesn't have to be this way. Maybe I
13:05 could keep u some elements but I feel
13:08 like again it communicates the point
13:10 that I was trying to make. Migrants who
13:12 perceive their English skills as
13:14 inadequate may may internalize feelings
13:17 of inferiority reinforcing inclusion and margin
13:18 margin
13:22 marginalization whatever. Um let's see
13:24 what we can do here. The sentence is
13:26 good. I like it. I would normally like
13:28 it. But then remember and this is
13:32 probably the what will cause us problem.
13:34 There is another sentence that's again
13:36 very similar in length in structure even
13:37 though it's a different sentence. So
13:39 then it says therefore supporting
13:40 self-esteem is vital not only for
13:42 linguistic competence but also for
13:45 promoting migrants overall well-being
13:48 and inclusion almost every time it lists
13:50 things is almost every time is two
13:53 things uh divided by and so it's just too
13:54 too
13:56 similar. Uh the previous sentence was
13:58 also about something and something and
14:00 then there is this almost exact same
14:03 structure. So I want to to avoid that. I
14:06 want to again get the overall sense of
14:09 what it means and combine it maybe into
14:11 now something that's a little bit longer
14:16 maybe slightly uh clumsy even so um as
14:20 and also rever reverse the order so as
14:24 low self-perceived or selfassessed we
14:28 could say English skills skills are
14:32 believed remember about toning down
14:34 uh to lead to inferior iority complex to
14:36 lead to
14:39 inferiority complex and exclusion
14:42 marginalization is I mean arguably
14:45 slightly different but for now it will
14:48 do an exclusion so as it's believed to
14:50 lead to these these things is vital it
14:54 says we have to do it so it is crucial
14:56 to support
15:00 self esteem building to promote the
15:03 migrants linguistic
15:07 and social inclusion. Uh because that's
15:09 what it says, promoting their well-being
15:12 and inclusion and the language
15:15 linguistic competence. So, this should
15:19 do. Hopefully, it's not too uh short
15:21 because I think it has to be 80 words.
15:25 It Yeah, that's enough uh for this uh
15:27 for this tool that we were using. So now
15:39 this
15:43 0%. It did take a little bit of of work
15:46 and I cut out some sequences when I was
15:48 thinking. Uh but that's basically what
15:51 happens. That's what happens. So I think
15:54 my main message to you would
15:57 be read it. Try to understand what it
16:00 says and try to say it in your own
16:02 words. In fact, what I would normally
16:04 do, I would probably have it as separate
16:06 paragraphs. I try to literally forget
16:08 what it said exactly. I just want to
16:10 remember the meaning and then say it in
16:11 my own words, but it's very hard because
16:14 you're influenced by what you just saw.
16:16 And it does have good sentences, but
16:17 basically that's what happens. I hope
16:18 that you enjoyed the video and learned
16:20 something new today. If you did, please
16:22 like the video. This really helps
16:24 YouTube position my video and lets
16:26 YouTube know that it's a good video with
16:29 quality content. Uh, feel free to ask me
16:31 questions in the comments. Feel free to
16:32 share it with others and consider
