Content Moderation Services in 2026: How to Keep Your Platform Safe, Legal, and Genuinely Human

Con­tent mod­er­a­tion ser­vices are man­aged pro­grams of peo­ple, pol­i­cy, and tech­nol­o­gy that review user-gen­er­at­ed text, images, video, audio, and AI out­puts to remove harm­ful or non-com­pli­ant mate­r­i­al before it reach­es your audi­ence. In plain terms, they are how a plat­form decides what stays up and what comes down, at a speed and scale no in-house team can match alone. This guide explains what these ser­vices actu­al­ly cov­er, the main types and trade-offs, how to com­pare providers, and a prac­ti­cal check­list you can use this quar­ter. You will also get an orig­i­nal deci­sion matrix for match­ing the right approach to your risk and vol­ume.

At a glance

Ques­tion	Direct answer
What are con­tent mod­er­a­tion ser­vices?	Out­sourced or man­aged pro­grams that review and act on user con­tent using trained human review­ers, pol­i­cy, and automa­tion.
Why do they mat­ter in 2026?	The con­tent mod­er­a­tion mar­ket reached about USD 9.7 bil­lion in 2023 and is fore­cast to hit USD 22.8 bil­lion by 2030 (Grand View Research), dri­ven by ris­ing UGC vol­umes and new laws.
What con­tent types are cov­ered?	Text and chat, image mod­er­a­tion, video mod­er­a­tion, audio, and gen­er­a­tive AI prompts and respons­es.
Human, AI, or both?	Hybrid (human-in-the-loop) is wide­ly regard­ed as the most reli­able mod­el because automa­tion miss­es con­text that trained review­ers catch.
What does the law require?	The EU Dig­i­tal Ser­vices Act man­dates trans­paren­cy report­ing and can fine plat­forms up to 6% of glob­al annu­al turnover for non-com­pli­ance.
How do I choose a part­ner?	Score providers on pol­i­cy fit, qual­i­ty assur­ance, lan­guages, turn­around, com­pli­ance, and review­er well­be­ing (see the check­list below).

Table of contents

What are con­tent mod­er­a­tion ser­vices?

Why con­tent mod­er­a­tion mat­ters right now

The main types of con­tent mod­er­a­tion

UGC mod­er­a­tion: the vol­ume prob­lem

Image mod­er­a­tion and video mod­er­a­tion

AI, human, or hybrid: a com­par­i­son table

When each approach actu­al­ly wins

The Graveiens Mod­er­a­tion Fit Matrix

How to choose a con­tent mod­er­a­tion part­ner

A worked exam­ple: before and after

How good mod­er­a­tion is real­ly built

Fre­quent­ly asked ques­tions

About the authors

Con­clu­sion

Sources

What are content moderation services?

Con­tent mod­er­a­tion ser­vices are man­aged pro­grams that review user-gen­er­at­ed con­tent and enforce a platform’s rules across every for­mat users can post. That means catch­ing hate speech in a com­ment thread, nudi­ty in an uploaded pho­to, graph­ic vio­lence in a live stream, scam links in a chat, and unsafe answers pro­duced by a gen­er­a­tive mod­el.

A mod­ern ser­vice is not just a room of review­ers. It com­bines a writ­ten pol­i­cy tax­on­o­my, auto­mat­ed pre-fil­ters that triage the easy cas­es, trained human review­ers who judge the hard ones, and a qual­i­ty-assur­ance loop that mea­sures whether deci­sions are cor­rect and con­sis­tent. The human lay­er exists because most dam­ag­ing con­tent is not obvi­ous: it is the bor­der­line post, the cod­ed slur, the satir­i­cal image, or the reclaimed word that auto­mat­ed fil­ters rou­tine­ly get wrong.

Providers usu­al­ly deliv­er this as a ful­ly man­aged team, an on-demand review queue, or a co-man­aged mod­el where the plat­form sets pol­i­cy and the part­ner sup­plies review­ers, tool­ing, and report­ing. The best pro­grams treat mod­er­a­tion as a liv­ing sys­tem that is cal­i­brat­ed against real exam­ples, not a one-time set­up.

Why content moderation matters right now

Two forces have turned con­tent mod­er­a­tion from a cost cen­ter into a board-lev­el pri­or­i­ty: scale and law.

On scale, the num­bers are mov­ing fast. Accord­ing to Grand View Research, the con­tent mod­er­a­tion mar­ket was val­ued at rough­ly USD 9.7 bil­lion in 2023 and is pro­ject­ed to reach USD 22.8 bil­lion by 2030, a com­pound annu­al growth rate of about 13% from 2024 to 2030. North Amer­i­ca account­ed for 33.9% of glob­al rev­enue in 2023, while Asia Pacif­ic is expect­ed to grow the fastest. The image seg­ment held the largest rev­enue share in 2023, and video is fore­cast to grow fastest, at more than 14% annu­al­ly. That last point mat­ters because visu­al for­mats are the hard­est and most expen­sive to review well.

On law, the reg­u­la­to­ry floor has risen. The EU Dig­i­tal Ser­vices Act now requires online plat­forms to pub­lish trans­paren­cy reports on their mod­er­a­tion activ­i­ty, includ­ing how much con­tent they remove and how accu­rate their auto­mat­ed sys­tems are. Very large online plat­forms must report twice a year; oth­er ser­vices report annu­al­ly. Har­mo­nized data col­lec­tion under the Commission’s imple­ment­ing rules began on July 1, 2025. Enforce­ment has teeth: fines can reach up to 6% of a provider’s glob­al annu­al turnover, with peri­od­ic penal­ties of up to 5% of aver­age dai­ly world­wide turnover for con­tin­ued non-com­pli­ance, accord­ing to guid­ance sum­ma­riz­ing the DSA’s enforce­ment pro­vi­sions. Get­ting mod­er­a­tion wrong is no longer just a rep­u­ta­tion­al risk. It is a finan­cial and legal one. In short, UGC mod­er­a­tion has moved from a back-office sup­port func­tion to a core oper­a­tion with real legal stakes.

The main types of content moderation

Con­tent mod­er­a­tion is usu­al­ly orga­nized in two ways: by the for­mat being reviewed, and by when the review hap­pens.

By for­mat, the main types are text and chat mod­er­a­tion, image mod­er­a­tion, video mod­er­a­tion, audio and voice mod­er­a­tion, and mod­er­a­tion of gen­er­a­tive AI out­puts (prompts and mod­el respons­es). Each has its own fail­ure modes. Text hinges on lan­guage and con­text. Image mod­er­a­tion depends on rec­og­niz­ing unsafe visu­als across cul­tures. Video mod­er­a­tion adds time, because harm can appear in a sin­gle frame buried in an hour of footage.

By tim­ing, teams choose among sev­er­al work­flows:

• Pre-mod­er­a­tion, where con­tent is reviewed before it goes live. Safest, but it adds laten­cy and does not scale to high-veloc­i­ty feeds.
• Post-mod­er­a­tion, where con­tent pub­lish­es imme­di­ate­ly and is reviewed short­ly after. The com­mon default for social and com­mu­ni­ty plat­forms.
• Reac­tive mod­er­a­tion, which relies on user reports to sur­face prob­lems. Cheap, but slow and eas­i­ly gamed.
• Proac­tive mod­er­a­tion, where automa­tion and review­ers hunt for vio­la­tions before any­one reports them. Essen­tial for the most seri­ous harms.

Most mature plat­forms run a blend: proac­tive detec­tion for severe cat­e­gories, post-mod­er­a­tion for the long tail, and pre-mod­er­a­tion only for the high­est-risk sur­faces.

UGC moderation: the volume problem

UGC mod­er­a­tion is the dis­ci­pline of review­ing user-gen­er­at­ed con­tent at inter­net scale, where mil­lions of posts, uploads, and mes­sages arrive every day and each one is a poten­tial lia­bil­i­ty. The core chal­lenge of UGC mod­er­a­tion is not any sin­gle deci­sion. It is doing mil­lions of them quick­ly, con­sis­tent­ly, and in dozens of lan­guages with­out burn­ing out review­ers or drown­ing users in false removals.

Effec­tive UGC mod­er­a­tion starts with triage, then routes each item to the right review­er. Auto­mat­ed clas­si­fiers han­dle the clear cas­es (obvi­ous spam, known ille­gal imagery) so that human atten­tion is reserved for gen­uine judg­ment calls. It also depends on a pre­cise pol­i­cy tax­on­o­my, because review­ers can­not apply a rule that is vague. A well-run user-gen­er­at­ed con­tent review pro­gram mea­sures inter-review­er agree­ment, tracks appeals, and feeds hard cas­es back into train­ing so the whole sys­tem keeps improv­ing.

Because UGC spans every for­mat, strong pro­grams coor­di­nate text, image mod­er­a­tion, and video mod­er­a­tion under one pol­i­cy rather than treat­ing them as sep­a­rate silos. That coor­di­na­tion is what keeps a cod­ed cap­tion from slip­ping through just because the image looked clean. In prac­tice, mature UGC mod­er­a­tion is best thought of as one sys­tem with many inputs, not a stack of dis­con­nect­ed tools.

Image moderation and video moderation

Visu­al con­tent is where mod­er­a­tion gets gen­uine­ly hard, which is why the fastest mar­ket growth is here. Get­ting image mod­er­a­tion and video mod­er­a­tion right is now a core com­pe­ten­cy for any plat­form that hosts uploads, not an after­thought.

Image moderation

Image mod­er­a­tion reviews pho­tos, graph­ics, memes, and pro­file pic­tures for unsafe or non-com­pli­ant visu­als, from explic­it mate­r­i­al to graph­ic vio­lence to hate sym­bols. Auto­mat­ed vision mod­els are strong on clear cat­e­gories but weak on con­text: a med­ical pho­to, a piece of art, and an exploita­tive image can share visu­al fea­tures, and only a trained review­er reli­ably tells them apart. Good image mod­er­a­tion pairs fast auto­mat­ed scor­ing with human review of the ambigu­ous mid­dle, plus cul­tur­al­ly aware review­ers who under­stand that a sym­bol or ges­ture can be benign in one region and hate­ful in anoth­er. Done well, image mod­er­a­tion pro­tects users with­out eras­ing legit­i­mate art, jour­nal­ism, or med­ical con­tent.

Video moderation

Video mod­er­a­tion is the most demand­ing for­mat because harm is spread across time and sound. A sin­gle unsafe frame, a slur in the audio track, or a manip­u­lat­ed clip can vio­late pol­i­cy inside oth­er­wise ordi­nary footage. Effec­tive video mod­er­a­tion com­bines frame sam­pling, audio tran­scrip­tion and review, and review­ers who can judge intent across a sequence rather than a still. For live video, the bar is high­er still: deci­sions have to hap­pen in sec­onds, which makes a well-drilled review­er team and clear esca­la­tion paths non-nego­tiable. This is why video is both the fastest-grow­ing and the most cost­ly seg­ment to staff cor­rect­ly.

AI, human, or hybrid: a comparison table

The sin­gle most impor­tant choice a plat­form makes is who or what does the review­ing. Here is how the main mod­els com­pare across the fac­tors that decide qual­i­ty and cost.

Fac­tor	AI-only automa­tion	Human-only review	Hybrid (human-in-the-loop)	Crowd­sourced / user reports
Speed at scale	Excel­lent	Poor	Very good	Mod­er­ate
Con­text accu­ra­cy	Weak on bor­der­line cas­es	Strong	Strong	Incon­sis­tent
Cost per item	Low­est	High­est	Mod­er­ate	Low
Mul­ti­lin­gual nuance	Lim­it­ed	Strong	Strong	Vari­able
Han­dles new harms	Slow to adapt	Fast	Fast	Slow
Auditabil­i­ty for reg­u­la­tors	Depends on log­ging	High	High	Low
Review­er well­be­ing risk	None	High expo­sure	Man­aged	Low
Best for	High-vol­ume clear cas­es	Small, sen­si­tive queues	Most real plat­forms	Sup­ple­men­tal sig­nal only

The pat­tern is con­sis­tent across inde­pen­dent research: automa­tion is fast and cheap but strug­gles with con­text, while humans are accu­rate but do not scale. A 2025 study in Nature Human Behav­iour found that auto­mat­ed sys­tems fre­quent­ly mis­judge con­tent because they can­not read con­text, some­times flag­ging harm­less posts, and that mul­ti­modal mod­els paired with human judg­ment pro­duce more con­text-sen­si­tive deci­sions. Ana­lysts at New Amer­i­ca have doc­u­ment­ed sim­i­lar lim­its in auto­mat­ed tools for years.

When each approach actually wins

The hon­est answer is that no sin­gle mod­el is best for every­one, and any­one who says oth­er­wise is sell­ing some­thing.

AI-only mod­er­a­tion wins when vol­ume is enor­mous and the cat­e­gories are unam­bigu­ous, such as fil­ter­ing known ille­gal imagery by hash-match­ing or block­ing obvi­ous spam. It fails when nuance mat­ters, because a clas­si­fi­er can­not tell satire from sin­cer­i­ty or reclaimed lan­guage from an attack.

Human-only review wins for small, high-stakes queues where every deci­sion is sen­si­tive, such as appeals, esca­la­tions, or trust-and-safe­ty inves­ti­ga­tions. It does not scale to a fire­hose of dai­ly uploads with­out unsus­tain­able cost and review­er strain.

Hybrid human-in-the-loop mod­er­a­tion wins for the vast major­i­ty of real plat­forms, because it puts automa­tion on the easy 80% and skilled review­ers on the 20% that car­ries almost all the risk. This is the mod­el most mod­er­a­tion pro­grams are built around today, and for good rea­son.

Crowd­sourced sig­nals such as user reports are use­ful as an input, nev­er as the sys­tem of record. They sur­face prob­lems the plat­form missed, but they are slow, eas­i­ly manip­u­lat­ed, and can­not be audit­ed against a regulator’s stan­dard.

The prac­ti­cal take­away is that the right design is usu­al­ly a lay­ered one: automa­tion for triage, humans for judg­ment, and clear rules for when a case jumps from one to the oth­er.

The Graveiens Moderation Fit Matrix

To make the choice repeat­able rather than a guess, we use a sim­ple deci­sion aid we call the Graveiens Mod­er­a­tion Fit Matrix. It maps two vari­ables every plat­form can esti­mate: con­tent risk (how much harm a missed item can cause) against con­tent vol­ume (how much arrives per day). The cell tells you where to invest first.

	Low vol­ume	High vol­ume
Low risk	Automa­tion with light spot-checks by humans; keep costs lean.	Automa­tion-led triage with sam­pled human QA; scale the clas­si­fiers.
High risk	Human-led review with automa­tion as an assis­tant; accu­ra­cy over speed.	Full human-in-the-loop pro­gram: proac­tive detec­tion, trained review­ers, four-stage QA, and 24/7 esca­la­tion.

The bot­tom-right quad­rant, high risk and high vol­ume, is where most large social, gam­ing, mar­ket­place, and gen­er­a­tive AI plat­forms actu­al­ly sit, and it is where man­aged con­tent mod­er­a­tion ser­vices earn their keep. It is also the quad­rant where dis­ci­plined UGC mod­er­a­tion stops being option­al and becomes a con­di­tion of stay­ing online. The matrix is delib­er­ate­ly sim­ple so any prod­uct or trust-and-safe­ty lead can place their plat­form in it dur­ing a sin­gle meet­ing and leave with a defen­si­ble start­ing point.

How to choose a content moderation partner

Use this check­list to eval­u­ate any UGC mod­er­a­tion provider. Score each item, because the gaps are where pro­grams fail.

1. Pol­i­cy fit. Can they adapt to your writ­ten tax­on­o­my and edge cas­es, or do they force a gener­ic rule­book onto your plat­form?

2. Qual­i­ty assur­ance. Ask for their QA mod­el. A seri­ous part­ner runs mul­ti-stage review, cal­i­brates against a gold-stan­dard test set, and reports accu­ra­cy after QA, not before.

3. Lan­guage and cul­ture. Con­firm real cov­er­age in the lan­guages and regions your users actu­al­ly post in, with review­ers who under­stand local con­text.

4. Turn­around and cov­er­age. Match their speed and hours to your risk. Live and high-sever­i­ty sur­faces need fast, around-the-clock response.

5. Com­pli­ance and auditabil­i­ty. Ver­i­fy they can pro­duce the logs and state­ments of rea­sons that laws like the DSA require, and ask about rel­e­vant cer­ti­fi­ca­tions such as ISO 9001:2017.

6. Review­er well­be­ing. Sus­tain­able pro­grams lim­it expo­sure, rotate review­ers, and pro­vide sup­port. This is both an eth­i­cal duty and a qual­i­ty safe­guard, because burned-out review­ers make worse deci­sions.

7. Com­mer­cial fair­ness. Look for mod­els that align pay­ment with accept­ed qual­i­ty, such as pay-on-approval pilots, rather than pay­ing for vol­ume regard­less of accu­ra­cy.

If a provider can­not answer these clear­ly, treat that as your answer.

A worked example: before and after

Con­sid­er a fast-grow­ing mar­ket­place app flood­ed with new list­ings and mes­sages.

Before: the team relied on an auto­mat­ed fil­ter plus user reports. The fil­ter blocked obvi­ous pro­fan­i­ty but missed scam list­ings writ­ten in care­ful lan­guage, and harm­ful items stayed live for hours until enough users com­plained. Appeals piled up, good sell­ers were wrong­ly removed, and the team could not pro­duce a clean audit trail when reg­u­la­tors asked.

After: the plat­form lay­ered a hybrid UGC mod­er­a­tion pro­gram on top. Automa­tion triaged the clear cas­es, trained review­ers han­dled the bor­der­line list­ings and mes­sages against a tight­ened tax­on­o­my, and a four-stage QA loop mea­sured accu­ra­cy. Scam list­ings were caught proac­tive­ly, false removals of legit­i­mate sell­ers dropped because a human made the final call, and every action pro­duced a logged rea­son ready for trans­paren­cy report­ing. The les­son is not that automa­tion failed. It is that automa­tion with­out human judg­ment left exact­ly the gaps that cost trust and invit­ed legal risk.

How good moderation is really built

Reli­able UGC mod­er­a­tion is a data-and-peo­ple prob­lem before it is a soft­ware prob­lem. The pro­grams that hold up under audit share a few traits: a pre­cise, ver­sioned pol­i­cy tax­on­o­my; automa­tion tuned to triage rather than to make final calls; trained mul­ti­lin­gual review­ers who own the judg­ment lay­er; and a mea­sur­able qual­i­ty-assur­ance work­flow that treats accu­ra­cy as a num­ber you improve, not a claim you make.

This is exact­ly where a spe­cial­ist part­ner adds val­ue. At Graveiens AI, our human-in-the-loop con­tent mod­er­a­tion pro­grams com­bine a cus­tomized pol­i­cy tax­on­o­my, trained review­ers across 25+ lan­guages, and a four-stage QA loop, with ISO 9001:2017 cer­ti­fied process­es and a pay-on-approval option for pilots so you only pay for review that meets the agreed bar. The same rig­or that pow­ers our data anno­ta­tion and data col­lec­tion work, where label­ing accu­ra­cy is every­thing, is what makes mod­er­a­tion defen­si­ble rather than best-effort.

Mod­er­a­tion is also expand­ing into new modal­i­ties. Gen­er­a­tive sys­tems now need their out­puts reviewed, which con­nects mod­er­a­tion direct­ly to LLM eval­u­a­tion and red-team­ing. And as AI moves into robot­ics and embod­ied sys­tems trained on first-per­son footage, the data behind those mod­els, such as con­sent-backed ego­cen­tric video cap­tured in real envi­ron­ments, needs the same care­ful review and gov­er­nance that mature mod­er­a­tion pro­grams already prac­tice. The com­mon thread is trust­wor­thy human judg­ment applied to messy, real-world con­tent.

For teams build­ing the train­ing pipelines under­neath all of this, our guide to AI train­ing data com­pa­nies is a use­ful com­pan­ion read, and if you are new to the fun­da­men­tals, start with what is train­ing data.

Frequently asked questions

What is the dif­fer­ence between con­tent mod­er­a­tion and con­tent mod­er­a­tion ser­vices?

Con­tent mod­er­a­tion is the act of review­ing and act­ing on user con­tent. Con­tent mod­er­a­tion ser­vices are the man­aged pro­grams, com­bin­ing peo­ple, pol­i­cy, and tech­nol­o­gy, that a plat­form buys or builds to do that work reli­ably at scale.

Are con­tent mod­er­a­tion ser­vices still need­ed if I use AI fil­ters?

Yes. AI fil­ters are excel­lent at triag­ing clear cas­es but weak on con­text, so they miss cod­ed lan­guage, satire, and cul­tur­al­ly spe­cif­ic harms. Inde­pen­dent research, includ­ing a 2025 Nature Human Behav­iour study, shows human-in-the-loop review pro­duces more con­text-accu­rate deci­sions than automa­tion alone.

How much do con­tent mod­er­a­tion ser­vices cost?

Pric­ing depends on vol­ume, for­mats, lan­guages, and turn­around. Visu­al for­mats like video mod­er­a­tion cost more than text because they take longer to review. Many providers offer usage-based or pay-on-approval pilot mod­els so you can val­i­date qual­i­ty before com­mit­ting.

What is UGC mod­er­a­tion?

UGC mod­er­a­tion is the review of user-gen­er­at­ed con­tent, such as posts, com­ments, images, and videos, at scale to enforce a platform’s rules and legal oblig­a­tions. It com­bines auto­mat­ed triage with trained human review­ers.

What is the dif­fer­ence between image mod­er­a­tion and video mod­er­a­tion?

Image mod­er­a­tion reviews still visu­als for unsafe con­tent. Video mod­er­a­tion is hard­er because harm can appear in a sin­gle frame or in the audio, so it adds frame sam­pling, audio review, and time-based judg­ment, and for live streams it must hap­pen in sec­onds.

Does the law require con­tent mod­er­a­tion?

In many mar­kets, effec­tive­ly yes. The EU Dig­i­tal Ser­vices Act requires trans­paren­cy report­ing on mod­er­a­tion and allows fines up to 6% of glob­al annu­al turnover, so doc­u­ment­ed, auditable mod­er­a­tion is a com­pli­ance require­ment, not an option.

Which is bet­ter, human or auto­mat­ed mod­er­a­tion?

For most plat­forms, a hybrid human-in-the-loop mod­el is best. Automa­tion han­dles the high-vol­ume clear cas­es and humans han­dle the bor­der­line judg­ments that car­ry the real risk. Pure automa­tion is faster but less accu­rate on con­text.

About the authors

This guide was writ­ten by the Graveiens AI Trust and Safe­ty Team, a group of mod­er­a­tion pro­gram man­agers, pol­i­cy spe­cial­ists, and mul­ti­lin­gual review­ers who design and run human-in-the-loop review at scale, and reviewed by a senior trust and safe­ty lead with over a decade of expe­ri­ence in online safe­ty and data oper­a­tions. Graveiens AI runs ISO 9001:2017 cer­ti­fied process­es and sup­ports mod­er­a­tion across 25+ lan­guages. Learn more on our About page.

Conclusion

Well-run con­tent mod­er­a­tion ser­vices are now the dif­fer­ence between a plat­form that earns trust and one that leaks it, and between pass­ing a regulator’s audit and pay­ing for fail­ing it. The mar­ket is grow­ing toward USD 22.8 bil­lion by 2030 for a rea­son: user con­tent keeps mul­ti­ply­ing, harms keep evolv­ing, and the law keeps tight­en­ing. Automa­tion alone can­not car­ry that weight, and human-only review can­not scale to meet it, which is why the lay­ered, human-in-the-loop mod­el has become the sen­si­ble default. If you are ready to make your mod­er­a­tion accu­rate, mul­ti­lin­gual, and audit-ready, talk to the Graveiens AI team on our con­tact page and we will help you design a pro­gram that fits your risk and vol­ume.

Sources

• Grand View Research, Con­tent Mod­er­a­tion Ser­vices Mar­ket Size Report, 2024 to 2030. https://www.grandviewresearch.com/industry-analysis/content-moderation-services-market-report
• Euro­pean Com­mis­sion, Com­mis­sion har­monis­es trans­paren­cy report­ing rules under the Dig­i­tal Ser­vices Act. https://digital-strategy.ec.europa.eu/en/news/commission-harmonises-transparency-reporting-rules-under-digital-services-act
• Euro­pean Inter­ac­tive Dig­i­tal Adver­tis­ing Alliance, Enforce­ment and Penal­ties under the EU Dig­i­tal Ser­vices Act. https://edaa.eu/digital-services-act/enforcement-and-penalties/
• David­son, T., Nature Human Behav­iour (2025), Con­text-sen­si­tive hate speech eval­u­a­tions with mul­ti­modal large lan­guage mod­els. https://www.nature.com/articles/s41562-025–02363‑7
• New Amer­i­ca, The Lim­i­ta­tions of Auto­mat­ed Tools in Con­tent Mod­er­a­tion. https://www.newamerica.org/insights/everything-moderation-analysis-how-internet-platforms-are-using-artificial-intelligence-moderate-user-generated-content/the-limitations-of-automated-tools-in-content-moderation/